state
stringlengths 0
159k
| srcUpToTactic
stringlengths 387
167k
| nextTactic
stringlengths 3
9k
| declUpToTactic
stringlengths 22
11.5k
| declId
stringlengths 38
95
| decl
stringlengths 16
1.89k
| file_tag
stringlengths 17
73
|
|---|---|---|---|---|---|---|
case pos
ι : Type u
s : Finset ι
a b : ℝ≥0∞
p q : ℝ
hpq : Real.IsConjugateExponent p q
h : a = ⊤ ∨ b = ⊤
⊢ ⊤ = a ^ p * (ENNReal.ofReal p)⁻¹ + b ^ q * (ENNReal.ofReal q)⁻¹ | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
| cases' h with h h | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
| Mathlib.Analysis.MeanInequalities.303_0.4hD1oATDjTWuML9 | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q | Mathlib_Analysis_MeanInequalities |
case pos.inl
ι : Type u
s : Finset ι
a b : ℝ≥0∞
p q : ℝ
hpq : Real.IsConjugateExponent p q
h : a = ⊤
⊢ ⊤ = a ^ p * (ENNReal.ofReal p)⁻¹ + b ^ q * (ENNReal.ofReal q)⁻¹ | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> | rw [h] | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> | Mathlib.Analysis.MeanInequalities.303_0.4hD1oATDjTWuML9 | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q | Mathlib_Analysis_MeanInequalities |
case pos.inr
ι : Type u
s : Finset ι
a b : ℝ≥0∞
p q : ℝ
hpq : Real.IsConjugateExponent p q
h : b = ⊤
⊢ ⊤ = a ^ p * (ENNReal.ofReal p)⁻¹ + b ^ q * (ENNReal.ofReal q)⁻¹ | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> | rw [h] | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> | Mathlib.Analysis.MeanInequalities.303_0.4hD1oATDjTWuML9 | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q | Mathlib_Analysis_MeanInequalities |
case pos.inl
ι : Type u
s : Finset ι
a b : ℝ≥0∞
p q : ℝ
hpq : Real.IsConjugateExponent p q
h : a = ⊤
⊢ ⊤ = ⊤ ^ p * (ENNReal.ofReal p)⁻¹ + b ^ q * (ENNReal.ofReal q)⁻¹ | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> | simp [h, hpq.pos, hpq.symm.pos] | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> | Mathlib.Analysis.MeanInequalities.303_0.4hD1oATDjTWuML9 | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q | Mathlib_Analysis_MeanInequalities |
case pos.inr
ι : Type u
s : Finset ι
a b : ℝ≥0∞
p q : ℝ
hpq : Real.IsConjugateExponent p q
h : b = ⊤
⊢ ⊤ = a ^ p * (ENNReal.ofReal p)⁻¹ + ⊤ ^ q * (ENNReal.ofReal q)⁻¹ | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> | simp [h, hpq.pos, hpq.symm.pos] | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> | Mathlib.Analysis.MeanInequalities.303_0.4hD1oATDjTWuML9 | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q | Mathlib_Analysis_MeanInequalities |
case neg
ι : Type u
s : Finset ι
a b : ℝ≥0∞
p q : ℝ
hpq : Real.IsConjugateExponent p q
h : ¬(a = ⊤ ∨ b = ⊤)
⊢ a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
| push_neg at h | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
| Mathlib.Analysis.MeanInequalities.303_0.4hD1oATDjTWuML9 | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q | Mathlib_Analysis_MeanInequalities |
case neg
ι : Type u
s : Finset ι
a b : ℝ≥0∞
p q : ℝ
hpq : Real.IsConjugateExponent p q
h : a ≠ ⊤ ∧ b ≠ ⊤
⊢ a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
| rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe] | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
| Mathlib.Analysis.MeanInequalities.303_0.4hD1oATDjTWuML9 | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
a b : ℝ≥0∞
p q : ℝ
hpq : Real.IsConjugateExponent p q
h : a ≠ ⊤ ∧ b ≠ ⊤
⊢ Real.toNNReal p ≠ 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by | simp [hpq.pos] | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by | Mathlib.Analysis.MeanInequalities.303_0.4hD1oATDjTWuML9 | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
a b : ℝ≥0∞
p q : ℝ
hpq : Real.IsConjugateExponent p q
h : a ≠ ⊤ ∧ b ≠ ⊤
⊢ Real.toNNReal q ≠ 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by | simp [hpq.symm.pos] | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by | Mathlib.Analysis.MeanInequalities.303_0.4hD1oATDjTWuML9 | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q | Mathlib_Analysis_MeanInequalities |
case neg
ι : Type u
s : Finset ι
a b : ℝ≥0∞
p q : ℝ
hpq : Real.IsConjugateExponent p q
h : a ≠ ⊤ ∧ b ≠ ⊤
⊢ ENNReal.toNNReal a * ENNReal.toNNReal b ≤
ENNReal.toNNReal a ^ p / Real.toNNReal p + ENNReal.toNNReal b ^ q / Real.toNNReal q | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
| exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
| Mathlib.Analysis.MeanInequalities.303_0.4hD1oATDjTWuML9 | /-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : ∑ i in s, f i ^ p ≤ 1
hg : ∑ i in s, g i ^ q ≤ 1
⊢ ∑ i in s, f i * g i ≤ 1 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
| have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
| Mathlib.Analysis.MeanInequalities.330_0.4hD1oATDjTWuML9 | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : ∑ i in s, f i ^ p ≤ 1
hg : ∑ i in s, g i ^ q ≤ 1
hp_ne_zero : Real.toNNReal p ≠ 0
⊢ ∑ i in s, f i * g i ≤ 1 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
| have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
| Mathlib.Analysis.MeanInequalities.330_0.4hD1oATDjTWuML9 | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : ∑ i in s, f i ^ p ≤ 1
hg : ∑ i in s, g i ^ q ≤ 1
hp_ne_zero : Real.toNNReal p ≠ 0
hq_ne_zero : Real.toNNReal q ≠ 0
⊢ ∑ i in s, f i * g i ≤ 1 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
| calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
| Mathlib.Analysis.MeanInequalities.330_0.4hD1oATDjTWuML9 | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : ∑ i in s, f i ^ p ≤ 1
hg : ∑ i in s, g i ^ q ≤ 1
hp_ne_zero : Real.toNNReal p ≠ 0
hq_ne_zero : Real.toNNReal q ≠ 0
⊢ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) =
(∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
| rw [sum_add_distrib, sum_div, sum_div] | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
| Mathlib.Analysis.MeanInequalities.330_0.4hD1oATDjTWuML9 | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : ∑ i in s, f i ^ p ≤ 1
hg : ∑ i in s, g i ^ q ≤ 1
hp_ne_zero : Real.toNNReal p ≠ 0
hq_ne_zero : Real.toNNReal q ≠ 0
⊢ (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q ≤
1 / Real.toNNReal p + 1 / Real.toNNReal q | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
| refine' add_le_add _ _ | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
| Mathlib.Analysis.MeanInequalities.330_0.4hD1oATDjTWuML9 | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 | Mathlib_Analysis_MeanInequalities |
case refine'_1
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : ∑ i in s, f i ^ p ≤ 1
hg : ∑ i in s, g i ^ q ≤ 1
hp_ne_zero : Real.toNNReal p ≠ 0
hq_ne_zero : Real.toNNReal q ≠ 0
⊢ (∑ i in s, f i ^ p) / Real.toNNReal p ≤ 1 / Real.toNNReal p | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· | rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero] | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· | Mathlib.Analysis.MeanInequalities.330_0.4hD1oATDjTWuML9 | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 | Mathlib_Analysis_MeanInequalities |
case refine'_2
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : ∑ i in s, f i ^ p ≤ 1
hg : ∑ i in s, g i ^ q ≤ 1
hp_ne_zero : Real.toNNReal p ≠ 0
hq_ne_zero : Real.toNNReal q ≠ 0
⊢ (∑ i in s, g i ^ q) / Real.toNNReal q ≤ 1 / Real.toNNReal q | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· | rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero] | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· | Mathlib.Analysis.MeanInequalities.330_0.4hD1oATDjTWuML9 | private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : ∑ i in s, f i ^ p = 0
⊢ ∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
| simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero] | private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
| Mathlib.Analysis.MeanInequalities.346_0.4hD1oATDjTWuML9 | private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : ∑ i in s, f i ^ p = 0
⊢ ∀ x ∈ s, f x = 0 ∨ g x = 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
| intro i his | private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
| Mathlib.Analysis.MeanInequalities.346_0.4hD1oATDjTWuML9 | private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : ∑ i in s, f i ^ p = 0
i : ι
his : i ∈ s
⊢ f i = 0 ∨ g i = 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
| left | private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
| Mathlib.Analysis.MeanInequalities.346_0.4hD1oATDjTWuML9 | private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case h
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : ∑ i in s, f i ^ p = 0
i : ι
his : i ∈ s
⊢ f i = 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
| rw [sum_eq_zero_iff] at hf | private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
| Mathlib.Analysis.MeanInequalities.346_0.4hD1oATDjTWuML9 | private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case h
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : ∀ x ∈ s, f x ^ p = 0
i : ι
his : i ∈ s
⊢ f i = 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
| exact (rpow_eq_zero_iff.mp (hf i his)).left | private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
| Mathlib.Analysis.MeanInequalities.346_0.4hD1oATDjTWuML9 | private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
⊢ ∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
| by_cases hF_zero : ∑ i in s, f i ^ p = 0 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
| Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case pos
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : ∑ i in s, f i ^ p = 0
⊢ ∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· | exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· | Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case neg
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
⊢ ∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
| by_cases hG_zero : ∑ i in s, g i ^ q = 0 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
| Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case pos
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : ∑ i in s, g i ^ q = 0
⊢ ∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· | calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _ | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· | Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : ∑ i in s, g i ^ q = 0
⊢ ∑ i in s, f i * g i = ∑ i in s, g i * f i | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
| congr with i | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
| Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case e_f.h.a
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : ∑ i in s, g i ^ q = 0
i : ι
⊢ ↑(f i * g i) = ↑(g i * f i) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
| rw [mul_comm] | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
| Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case neg
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : · i in s, g i ^ q = 0
⊢ ∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
| let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p) | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
| Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case neg
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : · i in s, g i ^ q = 0
f' : ι → ℝ≥0 := fun i => f i / (∑ i in s, f i ^ p) ^ (1 / p)
⊢ ∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
| let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q) | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
| Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case neg
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : · i in s, g i ^ q = 0
f' : ι → ℝ≥0 := fun i => f i / (∑ i in s, f i ^ p) ^ (1 / p)
g' : ι → ℝ≥0 := fun i => g i / (∑ i in s, g i ^ q) ^ (1 / q)
⊢ ∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
| suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
| Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : · i in s, g i ^ q = 0
f' : ι → ℝ≥0 := fun i => f i / (∑ i in s, f i ^ p) ^ (1 / p)
g' : ι → ℝ≥0 := fun i => g i / (∑ i in s, g i ^ q) ^ (1 / q)
this : ∑ i in s, f' i * g' i ≤ 1
⊢ ∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
| simp_rw [div_mul_div_comm, ← sum_div] at this | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
| Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : · i in s, g i ^ q = 0
f' : ι → ℝ≥0 := fun i => f i / (∑ i in s, f i ^ p) ^ (1 / p)
g' : ι → ℝ≥0 := fun i => g i / (∑ i in s, g i ^ q) ^ (1 / q)
this : (∑ i in s, f i * g i) / ((∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q)) ≤ 1
⊢ ∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
| rwa [div_le_iff, one_mul] at this | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
| Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : · i in s, g i ^ q = 0
f' : ι → ℝ≥0 := fun i => f i / (∑ i in s, f i ^ p) ^ (1 / p)
g' : ι → ℝ≥0 := fun i => g i / (∑ i in s, g i ^ q) ^ (1 / q)
this : (∑ i in s, f i * g i) / ((∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q)) ≤ 1
⊢ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) ≠ 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
| refine' mul_ne_zero _ _ | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
| Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case refine'_1
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : · i in s, g i ^ q = 0
f' : ι → ℝ≥0 := fun i => f i / (∑ i in s, f i ^ p) ^ (1 / p)
g' : ι → ℝ≥0 := fun i => g i / (∑ i in s, g i ^ q) ^ (1 / q)
this : (∑ i in s, f i * g i) / ((∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q)) ≤ 1
⊢ (∑ i in s, f i ^ p) ^ (1 / p) ≠ 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· | rw [Ne.def, rpow_eq_zero_iff, not_and_or] | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· | Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case refine'_1
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : · i in s, g i ^ q = 0
f' : ι → ℝ≥0 := fun i => f i / (∑ i in s, f i ^ p) ^ (1 / p)
g' : ι → ℝ≥0 := fun i => g i / (∑ i in s, g i ^ q) ^ (1 / q)
this : (∑ i in s, f i * g i) / ((∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q)) ≤ 1
⊢ ¬∑ i in s, f i ^ p = 0 ∨ ¬1 / p ≠ 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
| exact Or.inl hF_zero | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
| Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case refine'_2
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : · i in s, g i ^ q = 0
f' : ι → ℝ≥0 := fun i => f i / (∑ i in s, f i ^ p) ^ (1 / p)
g' : ι → ℝ≥0 := fun i => g i / (∑ i in s, g i ^ q) ^ (1 / q)
this : (∑ i in s, f i * g i) / ((∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q)) ≤ 1
⊢ (∑ i in s, g i ^ q) ^ (1 / q) ≠ 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· | rw [Ne.def, rpow_eq_zero_iff, not_and_or] | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· | Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case refine'_2
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : · i in s, g i ^ q = 0
f' : ι → ℝ≥0 := fun i => f i / (∑ i in s, f i ^ p) ^ (1 / p)
g' : ι → ℝ≥0 := fun i => g i / (∑ i in s, g i ^ q) ^ (1 / q)
this : (∑ i in s, f i * g i) / ((∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q)) ≤ 1
⊢ ¬∑ i in s, g i ^ q = 0 ∨ ¬1 / q ≠ 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
| exact Or.inl hG_zero | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
| Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case neg
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : · i in s, g i ^ q = 0
f' : ι → ℝ≥0 := fun i => f i / (∑ i in s, f i ^ p) ^ (1 / p)
g' : ι → ℝ≥0 := fun i => g i / (∑ i in s, g i ^ q) ^ (1 / q)
⊢ ∑ i in s, f' i * g' i ≤ 1 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
| refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _) | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
| Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case neg.refine'_1
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : · i in s, g i ^ q = 0
f' : ι → ℝ≥0 := fun i => f i / (∑ i in s, f i ^ p) ^ (1 / p)
g' : ι → ℝ≥0 := fun i => g i / (∑ i in s, g i ^ q) ^ (1 / q)
⊢ ∑ i in s, f' i ^ p = 1 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· | simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero] | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· | Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case neg.refine'_2
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hF_zero : · i in s, f i ^ p = 0
hG_zero : · i in s, g i ^ q = 0
f' : ι → ℝ≥0 := fun i => f i / (∑ i in s, f i ^ p) ^ (1 / p)
g' : ι → ℝ≥0 := fun i => g i / (∑ i in s, g i ^ q) ^ (1 / q)
⊢ ∑ i in s, g' i ^ q = 1 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· | simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero] | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· | Mathlib.Analysis.MeanInequalities.356_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ q
⊢ (Summable fun i => f i * g i) ∧
∑' (i : ι), f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
| have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
| Mathlib.Analysis.MeanInequalities.388_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ q
⊢ ∀ (s : Finset ι), ∑ i in s, f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
| intro s | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
| Mathlib.Analysis.MeanInequalities.388_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s✝ : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ q
s : Finset ι
⊢ ∑ i in s, f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
| refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le) | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
| Mathlib.Analysis.MeanInequalities.388_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case refine'_1
ι : Type u
s✝ : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ q
s : Finset ι
⊢ (∑ i in s, f i ^ p) ^ (1 / p) ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· | rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)] | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· | Mathlib.Analysis.MeanInequalities.388_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case refine'_1
ι : Type u
s✝ : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ q
s : Finset ι
⊢ ∑ i in s, f i ^ p ≤ ∑' (i : ι), f i ^ p | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
| exact sum_le_tsum _ (fun _ _ => zero_le _) hf | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
| Mathlib.Analysis.MeanInequalities.388_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case refine'_2
ι : Type u
s✝ : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ q
s : Finset ι
⊢ (∑ i in s, g i ^ q) ^ (1 / q) ≤ (∑' (i : ι), g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· | rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)] | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· | Mathlib.Analysis.MeanInequalities.388_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case refine'_2
ι : Type u
s✝ : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ q
s : Finset ι
⊢ ∑ i in s, g i ^ q ≤ ∑' (i : ι), g i ^ q | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
| exact sum_le_tsum _ (fun _ _ => zero_le _) hg | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
| Mathlib.Analysis.MeanInequalities.388_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ q
H₁ : ∀ (s : Finset ι), ∑ i in s, f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q)
⊢ (Summable fun i => f i * g i) ∧
∑' (i : ι), f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
| have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
| Mathlib.Analysis.MeanInequalities.388_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ q
H₁ : ∀ (s : Finset ι), ∑ i in s, f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q)
⊢ BddAbove (Set.range fun s => ∑ i in s, f i * g i) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
| refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩ | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
| Mathlib.Analysis.MeanInequalities.388_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ q
H₁ : ∀ (s : Finset ι), ∑ i in s, f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q)
⊢ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q) ∈
upperBounds (Set.range fun s => ∑ i in s, f i * g i) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
| rintro a ⟨s, rfl⟩ | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
| Mathlib.Analysis.MeanInequalities.388_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
case intro
ι : Type u
s✝ : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ q
H₁ : ∀ (s : Finset ι), ∑ i in s, f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q)
s : Finset ι
⊢ (fun s => ∑ i in s, f i * g i) s ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
| exact H₁ s | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
| Mathlib.Analysis.MeanInequalities.388_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ q
H₁ : ∀ (s : Finset ι), ∑ i in s, f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q)
bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i)
⊢ (Summable fun i => f i * g i) ∧
∑' (i : ι), f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
| have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
| Mathlib.Analysis.MeanInequalities.388_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ q
H₁ : ∀ (s : Finset ι), ∑ i in s, f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q)
bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i)
H₂ : Summable fun i => f i * g i
⊢ (Summable fun i => f i * g i) ∧
∑' (i : ι), f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
| exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩ | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
| Mathlib.Analysis.MeanInequalities.388_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
A B : ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : HasSum (fun i => f i ^ p) (A ^ p)
hg : HasSum (fun i => g i ^ q) (B ^ q)
⊢ ∃ C ≤ A * B, HasSum (fun i => f i * g i) C | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
| obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
| Mathlib.Analysis.MeanInequalities.424_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C | Mathlib_Analysis_MeanInequalities |
case intro
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
A B : ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : HasSum (fun i => f i ^ p) (A ^ p)
hg : HasSum (fun i => g i ^ q) (B ^ q)
H₁ : Summable fun i => f i * g i
H₂ : ∑' (i : ι), f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q)
⊢ ∃ C ≤ A * B, HasSum (fun i => f i * g i) C | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
| have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero] | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
| Mathlib.Analysis.MeanInequalities.424_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
A B : ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : HasSum (fun i => f i ^ p) (A ^ p)
hg : HasSum (fun i => g i ^ q) (B ^ q)
H₁ : Summable fun i => f i * g i
H₂ : ∑' (i : ι), f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q)
⊢ A = (∑' (i : ι), f i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by | rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero] | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by | Mathlib.Analysis.MeanInequalities.424_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C | Mathlib_Analysis_MeanInequalities |
case intro
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
A B : ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : HasSum (fun i => f i ^ p) (A ^ p)
hg : HasSum (fun i => g i ^ q) (B ^ q)
H₁ : Summable fun i => f i * g i
H₂ : ∑' (i : ι), f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q)
hA : A = (∑' (i : ι), f i ^ p) ^ (1 / p)
⊢ ∃ C ≤ A * B, HasSum (fun i => f i * g i) C | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
| have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero] | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
| Mathlib.Analysis.MeanInequalities.424_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
A B : ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : HasSum (fun i => f i ^ p) (A ^ p)
hg : HasSum (fun i => g i ^ q) (B ^ q)
H₁ : Summable fun i => f i * g i
H₂ : ∑' (i : ι), f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q)
hA : A = (∑' (i : ι), f i ^ p) ^ (1 / p)
⊢ B = (∑' (i : ι), g i ^ q) ^ (1 / q) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
| rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero] | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
| Mathlib.Analysis.MeanInequalities.424_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C | Mathlib_Analysis_MeanInequalities |
case intro
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
A B : ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : HasSum (fun i => f i ^ p) (A ^ p)
hg : HasSum (fun i => g i ^ q) (B ^ q)
H₁ : Summable fun i => f i * g i
H₂ : ∑' (i : ι), f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q)
hA : A = (∑' (i : ι), f i ^ p) ^ (1 / p)
hB : B = (∑' (i : ι), g i ^ q) ^ (1 / q)
⊢ ∃ C ≤ A * B, HasSum (fun i => f i * g i) C | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
| refine' ⟨∑' i, f i * g i, _, _⟩ | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
| Mathlib.Analysis.MeanInequalities.424_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C | Mathlib_Analysis_MeanInequalities |
case intro.refine'_1
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
A B : ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : HasSum (fun i => f i ^ p) (A ^ p)
hg : HasSum (fun i => g i ^ q) (B ^ q)
H₁ : Summable fun i => f i * g i
H₂ : ∑' (i : ι), f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q)
hA : A = (∑' (i : ι), f i ^ p) ^ (1 / p)
hB : B = (∑' (i : ι), g i ^ q) ^ (1 / q)
⊢ ∑' (i : ι), f i * g i ≤ A * B | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· | simpa [hA, hB] using H₂ | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· | Mathlib.Analysis.MeanInequalities.424_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C | Mathlib_Analysis_MeanInequalities |
case intro.refine'_2
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
A B : ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : HasSum (fun i => f i ^ p) (A ^ p)
hg : HasSum (fun i => g i ^ q) (B ^ q)
H₁ : Summable fun i => f i * g i
H₂ : ∑' (i : ι), f i * g i ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) * (∑' (i : ι), g i ^ q) ^ (1 / q)
hA : A = (∑' (i : ι), f i ^ p) ^ (1 / p)
hB : B = (∑' (i : ι), g i ^ q) ^ (1 / q)
⊢ HasSum (fun i => f i * g i) (∑' (i : ι), f i * g i) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· | simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· | Mathlib.Analysis.MeanInequalities.424_0.4hD1oATDjTWuML9 | /-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p : ℝ
hp : 1 ≤ p
⊢ (∑ i in s, f i) ^ p ≤ ↑(card s) ^ (p - 1) * ∑ i in s, f i ^ p | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
| cases' eq_or_lt_of_le hp with hp hp | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
| Mathlib.Analysis.MeanInequalities.440_0.4hD1oATDjTWuML9 | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p | Mathlib_Analysis_MeanInequalities |
case inl
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 = p
⊢ (∑ i in s, f i) ^ p ≤ ↑(card s) ^ (p - 1) * ∑ i in s, f i ^ p | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· | simp [← hp] | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· | Mathlib.Analysis.MeanInequalities.440_0.4hD1oATDjTWuML9 | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p | Mathlib_Analysis_MeanInequalities |
case inr
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
⊢ (∑ i in s, f i) ^ p ≤ ↑(card s) ^ (p - 1) * ∑ i in s, f i ^ p | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
| let q : ℝ := p / (p - 1) | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
| Mathlib.Analysis.MeanInequalities.440_0.4hD1oATDjTWuML9 | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p | Mathlib_Analysis_MeanInequalities |
case inr
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
q : ℝ := p / (p - 1)
⊢ (∑ i in s, f i) ^ p ≤ ↑(card s) ^ (p - 1) * ∑ i in s, f i ^ p | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
| have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp] | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
| Mathlib.Analysis.MeanInequalities.440_0.4hD1oATDjTWuML9 | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
q : ℝ := p / (p - 1)
⊢ Real.IsConjugateExponent p q | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by | rw [Real.isConjugateExponent_iff hp] | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by | Mathlib.Analysis.MeanInequalities.440_0.4hD1oATDjTWuML9 | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p | Mathlib_Analysis_MeanInequalities |
case inr
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
q : ℝ := p / (p - 1)
hpq : Real.IsConjugateExponent p q
⊢ (∑ i in s, f i) ^ p ≤ ↑(card s) ^ (p - 1) * ∑ i in s, f i ^ p | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
| have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
| Mathlib.Analysis.MeanInequalities.440_0.4hD1oATDjTWuML9 | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p | Mathlib_Analysis_MeanInequalities |
case inr
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
q : ℝ := p / (p - 1)
hpq : Real.IsConjugateExponent p q
hp₁ : 1 / p * p = 1
⊢ (∑ i in s, f i) ^ p ≤ ↑(card s) ^ (p - 1) * ∑ i in s, f i ^ p | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
| have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
| Mathlib.Analysis.MeanInequalities.440_0.4hD1oATDjTWuML9 | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
q : ℝ := p / (p - 1)
hpq : Real.IsConjugateExponent p q
hp₁ : 1 / p * p = 1
⊢ 1 / q * p = p - 1 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
| rw [← hpq.div_conj_eq_sub_one] | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
| Mathlib.Analysis.MeanInequalities.440_0.4hD1oATDjTWuML9 | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
q : ℝ := p / (p - 1)
hpq : Real.IsConjugateExponent p q
hp₁ : 1 / p * p = 1
⊢ 1 / q * p = p / q | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
| ring | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
| Mathlib.Analysis.MeanInequalities.440_0.4hD1oATDjTWuML9 | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p | Mathlib_Analysis_MeanInequalities |
case inr
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
q : ℝ := p / (p - 1)
hpq : Real.IsConjugateExponent p q
hp₁ : 1 / p * p = 1
hq : 1 / q * p = p - 1
⊢ (∑ i in s, f i) ^ p ≤ ↑(card s) ^ (p - 1) * ∑ i in s, f i ^ p | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
| simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
| Mathlib.Analysis.MeanInequalities.440_0.4hD1oATDjTWuML9 | /-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
⊢ IsGreatest ((fun g => ∑ i in s, f i * g i) '' {g | ∑ i in s, g i ^ q ≤ 1}) ((∑ i in s, f i ^ p) ^ (1 / p)) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
| constructor | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
| Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
case left
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
⊢ (∑ i in s, f i ^ p) ^ (1 / p) ∈ (fun g => ∑ i in s, f i * g i) '' {g | ∑ i in s, g i ^ q ≤ 1} | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· | use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q) | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· | Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
case h
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
⊢ (fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)) ∈ {g | ∑ i in s, g i ^ q ≤ 1} ∧
((fun g => ∑ i in s, f i * g i) fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)) =
(∑ i in s, f i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
| by_cases hf : ∑ i in s, f i ^ p = 0 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
| Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
case pos
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : ∑ i in s, f i ^ p = 0
⊢ (fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)) ∈ {g | ∑ i in s, g i ^ q ≤ 1} ∧
((fun g => ∑ i in s, f i * g i) fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)) =
(∑ i in s, f i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· | simp [hf, hpq.ne_zero, hpq.symm.ne_zero] | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· | Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
case neg
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : · i in s, f i ^ p = 0
⊢ (fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)) ∈ {g | ∑ i in s, g i ^ q ≤ 1} ∧
((fun g => ∑ i in s, f i * g i) fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)) =
(∑ i in s, f i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· | have A : p + q - q ≠ 0 := by simp [hpq.ne_zero] | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· | Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : · i in s, f i ^ p = 0
⊢ p + q - q ≠ 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by | simp [hpq.ne_zero] | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by | Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
case neg
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : · i in s, f i ^ p = 0
A : p + q - q ≠ 0
⊢ (fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)) ∈ {g | ∑ i in s, g i ^ q ≤ 1} ∧
((fun g => ∑ i in s, f i * g i) fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)) =
(∑ i in s, f i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
| have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero] | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
| Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : · i in s, f i ^ p = 0
A : p + q - q ≠ 0
⊢ ∀ (y : ℝ≥0), y * y ^ p / y = y ^ p | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
| refine' fun y => mul_div_cancel_left_of_imp fun h => _ | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
| Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : · i in s, f i ^ p = 0
A : p + q - q ≠ 0
y : ℝ≥0
h : y = 0
⊢ y ^ p = 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
| simp [h, hpq.ne_zero] | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
| Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
case neg
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : · i in s, f i ^ p = 0
A : p + q - q ≠ 0
B : ∀ (y : ℝ≥0), y * y ^ p / y = y ^ p
⊢ (fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)) ∈ {g | ∑ i in s, g i ^ q ≤ 1} ∧
((fun g => ∑ i in s, f i * g i) fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)) =
(∑ i in s, f i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
| simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B] | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
| Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
case neg
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : · i in s, f i ^ p = 0
A : p + q - q ≠ 0
B : ∀ (y : ℝ≥0), y * y ^ p / y = y ^ p
⊢ (∑ i in s, f i ^ p) / (∑ i in s, f i ^ p) ^ (1 / q) = (∑ i in s, f i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
| rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one] | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
| Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
case neg
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
hf : · i in s, f i ^ p = 0
A : p + q - q ≠ 0
B : ∀ (y : ℝ≥0), y * y ^ p / y = y ^ p
⊢ (∑ i in s, f i ^ p) ^ (1 / q) ≠ 0 | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
| simpa [hpq.symm.ne_zero] using hf | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
| Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
case right
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
⊢ (∑ i in s, f i ^ p) ^ (1 / p) ∈ upperBounds ((fun g => ∑ i in s, f i * g i) '' {g | ∑ i in s, g i ^ q ≤ 1}) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· | rintro _ ⟨g, hg, rfl⟩ | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· | Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
case right.intro.intro
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
g : ι → ℝ≥0
hg : g ∈ {g | ∑ i in s, g i ^ q ≤ 1}
⊢ (fun g => ∑ i in s, f i * g i) g ≤ (∑ i in s, f i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
| apply le_trans (inner_le_Lp_mul_Lq s f g hpq) | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
| Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
case right.intro.intro
ι : Type u
s : Finset ι
f : ι → ℝ≥0
p q : ℝ
hpq : Real.IsConjugateExponent p q
g : ι → ℝ≥0
hg : g ∈ {g | ∑ i in s, g i ^ q ≤ 1}
⊢ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) ≤ (∑ i in s, f i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
| simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _ | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
| Mathlib.Analysis.MeanInequalities.458_0.4hD1oATDjTWuML9 | /-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p : ℝ
hp : 1 ≤ p
⊢ (∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤ (∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _
#align nnreal.is_greatest_Lp NNReal.isGreatest_Lp
/-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
| rcases eq_or_lt_of_le hp with (rfl | hp) | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
| Mathlib.Analysis.MeanInequalities.482_0.4hD1oATDjTWuML9 | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | Mathlib_Analysis_MeanInequalities |
case inl
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
hp : 1 ≤ 1
⊢ (∑ i in s, (f i + g i) ^ 1) ^ (1 / 1) ≤ (∑ i in s, f i ^ 1) ^ (1 / 1) + (∑ i in s, g i ^ 1) ^ (1 / 1) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _
#align nnreal.is_greatest_Lp NNReal.isGreatest_Lp
/-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· | simp [Finset.sum_add_distrib] | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· | Mathlib.Analysis.MeanInequalities.482_0.4hD1oATDjTWuML9 | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | Mathlib_Analysis_MeanInequalities |
case inr
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
⊢ (∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤ (∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _
#align nnreal.is_greatest_Lp NNReal.isGreatest_Lp
/-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
| have hpq := Real.isConjugateExponent_conjugateExponent hp | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
| Mathlib.Analysis.MeanInequalities.482_0.4hD1oATDjTWuML9 | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | Mathlib_Analysis_MeanInequalities |
case inr
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
hpq : Real.IsConjugateExponent p (Real.conjugateExponent p)
⊢ (∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤ (∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _
#align nnreal.is_greatest_Lp NNReal.isGreatest_Lp
/-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
| have := isGreatest_Lp s (f + g) hpq | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
| Mathlib.Analysis.MeanInequalities.482_0.4hD1oATDjTWuML9 | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | Mathlib_Analysis_MeanInequalities |
case inr
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
hpq : Real.IsConjugateExponent p (Real.conjugateExponent p)
this :
IsGreatest ((fun g_1 => ∑ i in s, (f + g) i * g_1 i) '' {g | ∑ i in s, g i ^ Real.conjugateExponent p ≤ 1})
((∑ i in s, (f + g) i ^ p) ^ (1 / p))
⊢ (∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤ (∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _
#align nnreal.is_greatest_Lp NNReal.isGreatest_Lp
/-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
have := isGreatest_Lp s (f + g) hpq
| simp only [Pi.add_apply, add_mul, sum_add_distrib] at this | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
have := isGreatest_Lp s (f + g) hpq
| Mathlib.Analysis.MeanInequalities.482_0.4hD1oATDjTWuML9 | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | Mathlib_Analysis_MeanInequalities |
case inr
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
hpq : Real.IsConjugateExponent p (Real.conjugateExponent p)
this :
IsGreatest
((fun a => ∑ x in s, f x * a x + ∑ x in s, g x * a x) '' {g | ∑ i in s, g i ^ Real.conjugateExponent p ≤ 1})
((∑ x in s, (f x + g x) ^ p) ^ (1 / p))
⊢ (∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤ (∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _
#align nnreal.is_greatest_Lp NNReal.isGreatest_Lp
/-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
have := isGreatest_Lp s (f + g) hpq
simp only [Pi.add_apply, add_mul, sum_add_distrib] at this
| rcases this.1 with ⟨φ, hφ, H⟩ | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
have := isGreatest_Lp s (f + g) hpq
simp only [Pi.add_apply, add_mul, sum_add_distrib] at this
| Mathlib.Analysis.MeanInequalities.482_0.4hD1oATDjTWuML9 | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | Mathlib_Analysis_MeanInequalities |
case inr.intro.intro
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
hpq : Real.IsConjugateExponent p (Real.conjugateExponent p)
this :
IsGreatest
((fun a => ∑ x in s, f x * a x + ∑ x in s, g x * a x) '' {g | ∑ i in s, g i ^ Real.conjugateExponent p ≤ 1})
((∑ x in s, (f x + g x) ^ p) ^ (1 / p))
φ : ι → ℝ≥0
hφ : φ ∈ {g | ∑ i in s, g i ^ Real.conjugateExponent p ≤ 1}
H : (fun a => ∑ x in s, f x * a x + ∑ x in s, g x * a x) φ = (∑ x in s, (f x + g x) ^ p) ^ (1 / p)
⊢ (∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤ (∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _
#align nnreal.is_greatest_Lp NNReal.isGreatest_Lp
/-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
have := isGreatest_Lp s (f + g) hpq
simp only [Pi.add_apply, add_mul, sum_add_distrib] at this
rcases this.1 with ⟨φ, hφ, H⟩
| rw [← H] | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
have := isGreatest_Lp s (f + g) hpq
simp only [Pi.add_apply, add_mul, sum_add_distrib] at this
rcases this.1 with ⟨φ, hφ, H⟩
| Mathlib.Analysis.MeanInequalities.482_0.4hD1oATDjTWuML9 | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | Mathlib_Analysis_MeanInequalities |
case inr.intro.intro
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p : ℝ
hp✝ : 1 ≤ p
hp : 1 < p
hpq : Real.IsConjugateExponent p (Real.conjugateExponent p)
this :
IsGreatest
((fun a => ∑ x in s, f x * a x + ∑ x in s, g x * a x) '' {g | ∑ i in s, g i ^ Real.conjugateExponent p ≤ 1})
((∑ x in s, (f x + g x) ^ p) ^ (1 / p))
φ : ι → ℝ≥0
hφ : φ ∈ {g | ∑ i in s, g i ^ Real.conjugateExponent p ≤ 1}
H : (fun a => ∑ x in s, f x * a x + ∑ x in s, g x * a x) φ = (∑ x in s, (f x + g x) ^ p) ^ (1 / p)
⊢ (fun a => ∑ x in s, f x * a x + ∑ x in s, g x * a x) φ ≤ (∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _
#align nnreal.is_greatest_Lp NNReal.isGreatest_Lp
/-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
have := isGreatest_Lp s (f + g) hpq
simp only [Pi.add_apply, add_mul, sum_add_distrib] at this
rcases this.1 with ⟨φ, hφ, H⟩
rw [← H]
| exact
add_le_add ((isGreatest_Lp s f hpq).2 ⟨φ, hφ, rfl⟩) ((isGreatest_Lp s g hpq).2 ⟨φ, hφ, rfl⟩) | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
have := isGreatest_Lp s (f + g) hpq
simp only [Pi.add_apply, add_mul, sum_add_distrib] at this
rcases this.1 with ⟨φ, hφ, H⟩
rw [← H]
| Mathlib.Analysis.MeanInequalities.482_0.4hD1oATDjTWuML9 | /-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p : ℝ
hp : 1 ≤ p
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ p
⊢ (Summable fun i => (f i + g i) ^ p) ∧
(∑' (i : ι), (f i + g i) ^ p) ^ (1 / p) ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) + (∑' (i : ι), g i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _
#align nnreal.is_greatest_Lp NNReal.isGreatest_Lp
/-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
have := isGreatest_Lp s (f + g) hpq
simp only [Pi.add_apply, add_mul, sum_add_distrib] at this
rcases this.1 with ⟨φ, hφ, H⟩
rw [← H]
exact
add_le_add ((isGreatest_Lp s f hpq).2 ⟨φ, hφ, rfl⟩) ((isGreatest_Lp s g hpq).2 ⟨φ, hφ, rfl⟩)
#align nnreal.Lp_add_le NNReal.Lp_add_le
/-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) := by
| have pos : 0 < p := lt_of_lt_of_le zero_lt_one hp | /-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) := by
| Mathlib.Analysis.MeanInequalities.499_0.4hD1oATDjTWuML9 | /-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p : ℝ
hp : 1 ≤ p
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ p
pos : 0 < p
⊢ (Summable fun i => (f i + g i) ^ p) ∧
(∑' (i : ι), (f i + g i) ^ p) ^ (1 / p) ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) + (∑' (i : ι), g i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _
#align nnreal.is_greatest_Lp NNReal.isGreatest_Lp
/-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
have := isGreatest_Lp s (f + g) hpq
simp only [Pi.add_apply, add_mul, sum_add_distrib] at this
rcases this.1 with ⟨φ, hφ, H⟩
rw [← H]
exact
add_le_add ((isGreatest_Lp s f hpq).2 ⟨φ, hφ, rfl⟩) ((isGreatest_Lp s g hpq).2 ⟨φ, hφ, rfl⟩)
#align nnreal.Lp_add_le NNReal.Lp_add_le
/-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) := by
have pos : 0 < p := lt_of_lt_of_le zero_lt_one hp
| have H₁ : ∀ s : Finset ι,
(∑ i in s, (f i + g i) ^ p) ≤
((∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p)) ^ p := by
intro s
rw [← NNReal.rpow_one_div_le_iff pos]
refine' le_trans (Lp_add_le s f g hp) (add_le_add _ _) <;>
rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr pos)] <;>
refine' sum_le_tsum _ (fun _ _ => zero_le _) _
exacts [hf, hg] | /-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) := by
have pos : 0 < p := lt_of_lt_of_le zero_lt_one hp
| Mathlib.Analysis.MeanInequalities.499_0.4hD1oATDjTWuML9 | /-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s : Finset ι
f g : ι → ℝ≥0
p : ℝ
hp : 1 ≤ p
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ p
pos : 0 < p
⊢ ∀ (s : Finset ι), ∑ i in s, (f i + g i) ^ p ≤ ((∑' (i : ι), f i ^ p) ^ (1 / p) + (∑' (i : ι), g i ^ p) ^ (1 / p)) ^ p | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _
#align nnreal.is_greatest_Lp NNReal.isGreatest_Lp
/-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
have := isGreatest_Lp s (f + g) hpq
simp only [Pi.add_apply, add_mul, sum_add_distrib] at this
rcases this.1 with ⟨φ, hφ, H⟩
rw [← H]
exact
add_le_add ((isGreatest_Lp s f hpq).2 ⟨φ, hφ, rfl⟩) ((isGreatest_Lp s g hpq).2 ⟨φ, hφ, rfl⟩)
#align nnreal.Lp_add_le NNReal.Lp_add_le
/-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) := by
have pos : 0 < p := lt_of_lt_of_le zero_lt_one hp
have H₁ : ∀ s : Finset ι,
(∑ i in s, (f i + g i) ^ p) ≤
((∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p)) ^ p := by
| intro s | /-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) := by
have pos : 0 < p := lt_of_lt_of_le zero_lt_one hp
have H₁ : ∀ s : Finset ι,
(∑ i in s, (f i + g i) ^ p) ≤
((∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p)) ^ p := by
| Mathlib.Analysis.MeanInequalities.499_0.4hD1oATDjTWuML9 | /-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s✝ : Finset ι
f g : ι → ℝ≥0
p : ℝ
hp : 1 ≤ p
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ p
pos : 0 < p
s : Finset ι
⊢ ∑ i in s, (f i + g i) ^ p ≤ ((∑' (i : ι), f i ^ p) ^ (1 / p) + (∑' (i : ι), g i ^ p) ^ (1 / p)) ^ p | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _
#align nnreal.is_greatest_Lp NNReal.isGreatest_Lp
/-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
have := isGreatest_Lp s (f + g) hpq
simp only [Pi.add_apply, add_mul, sum_add_distrib] at this
rcases this.1 with ⟨φ, hφ, H⟩
rw [← H]
exact
add_le_add ((isGreatest_Lp s f hpq).2 ⟨φ, hφ, rfl⟩) ((isGreatest_Lp s g hpq).2 ⟨φ, hφ, rfl⟩)
#align nnreal.Lp_add_le NNReal.Lp_add_le
/-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) := by
have pos : 0 < p := lt_of_lt_of_le zero_lt_one hp
have H₁ : ∀ s : Finset ι,
(∑ i in s, (f i + g i) ^ p) ≤
((∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p)) ^ p := by
intro s
| rw [← NNReal.rpow_one_div_le_iff pos] | /-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) := by
have pos : 0 < p := lt_of_lt_of_le zero_lt_one hp
have H₁ : ∀ s : Finset ι,
(∑ i in s, (f i + g i) ^ p) ≤
((∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p)) ^ p := by
intro s
| Mathlib.Analysis.MeanInequalities.499_0.4hD1oATDjTWuML9 | /-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) | Mathlib_Analysis_MeanInequalities |
ι : Type u
s✝ : Finset ι
f g : ι → ℝ≥0
p : ℝ
hp : 1 ≤ p
hf : Summable fun i => f i ^ p
hg : Summable fun i => g i ^ p
pos : 0 < p
s : Finset ι
⊢ (∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤ (∑' (i : ι), f i ^ p) ^ (1 / p) + (∑' (i : ι), g i ^ p) ^ (1 / p) | /-
Copyright (c) 2019 Yury Kudryashov. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Yury Kudryashov, Sébastien Gouëzel, Rémy Degenne
-/
import Mathlib.Analysis.Convex.Jensen
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.SpecialFunctions.Pow.NNReal
import Mathlib.Data.Real.ConjugateExponents
#align_import analysis.mean_inequalities from "leanprover-community/mathlib"@"8f9fea08977f7e450770933ee6abb20733b47c92"
/-!
# Mean value inequalities
In this file we prove several inequalities for finite sums, including AM-GM inequality,
Young's inequality, Hölder inequality, and Minkowski inequality. Versions for integrals of some of
these inequalities are available in `MeasureTheory.MeanInequalities`.
## Main theorems
### AM-GM inequality:
The inequality says that the geometric mean of a tuple of non-negative numbers is less than or equal
to their arithmetic mean. We prove the weighted version of this inequality: if $w$ and $z$
are two non-negative vectors and $\sum_{i\in s} w_i=1$, then
$$
\prod_{i\in s} z_i^{w_i} ≤ \sum_{i\in s} w_iz_i.
$$
The classical version is a special case of this inequality for $w_i=\frac{1}{n}$.
We prove a few versions of this inequality. Each of the following lemmas comes in two versions:
a version for real-valued non-negative functions is in the `Real` namespace, and a version for
`NNReal`-valued functions is in the `NNReal` namespace.
- `geom_mean_le_arith_mean_weighted` : weighted version for functions on `Finset`s;
- `geom_mean_le_arith_mean2_weighted` : weighted version for two numbers;
- `geom_mean_le_arith_mean3_weighted` : weighted version for three numbers;
- `geom_mean_le_arith_mean4_weighted` : weighted version for four numbers.
### Young's inequality
Young's inequality says that for non-negative numbers `a`, `b`, `p`, `q` such that
$\frac{1}{p}+\frac{1}{q}=1$ we have
$$
ab ≤ \frac{a^p}{p} + \frac{b^q}{q}.
$$
This inequality is a special case of the AM-GM inequality. It is then used to prove Hölder's
inequality (see below).
### Hölder's inequality
The inequality says that for two conjugate exponents `p` and `q` (i.e., for two positive numbers
such that $\frac{1}{p}+\frac{1}{q}=1$) and any two non-negative vectors their inner product is
less than or equal to the product of the $L_p$ norm of the first vector and the $L_q$ norm of the
second vector:
$$
\sum_{i\in s} a_ib_i ≤ \sqrt[p]{\sum_{i\in s} a_i^p}\sqrt[q]{\sum_{i\in s} b_i^q}.
$$
We give versions of this result in `ℝ`, `ℝ≥0` and `ℝ≥0∞`.
There are at least two short proofs of this inequality. In our proof we prenormalize both vectors,
then apply Young's inequality to each $a_ib_i$. Another possible proof would be to deduce this
inequality from the generalized mean inequality for well-chosen vectors and weights.
### Minkowski's inequality
The inequality says that for `p ≥ 1` the function
$$
\|a\|_p=\sqrt[p]{\sum_{i\in s} a_i^p}
$$
satisfies the triangle inequality $\|a+b\|_p\le \|a\|_p+\|b\|_p$.
We give versions of this result in `Real`, `ℝ≥0` and `ℝ≥0∞`.
We deduce this inequality from Hölder's inequality. Namely, Hölder inequality implies that $\|a\|_p$
is the maximum of the inner product $\sum_{i\in s}a_ib_i$ over `b` such that $\|b\|_q\le 1$. Now
Minkowski's inequality follows from the fact that the maximum value of the sum of two functions is
less than or equal to the sum of the maximum values of the summands.
## TODO
- each inequality `A ≤ B` should come with a theorem `A = B ↔ _`; one of the ways to prove them
is to define `StrictConvexOn` functions.
- generalized mean inequality with any `p ≤ q`, including negative numbers;
- prove that the power mean tends to the geometric mean as the exponent tends to zero.
-/
universe u v
open Finset Classical BigOperators NNReal ENNReal
set_option linter.uppercaseLean3 false
noncomputable section
variable {ι : Type u} (s : Finset ι)
section GeomMeanLEArithMean
/-! ### AM-GM inequality -/
namespace Real
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**, weighted
version for real-valued nonnegative functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) :
∏ i in s, z i ^ w i ≤ ∑ i in s, w i * z i := by
-- If some number `z i` equals zero and has non-zero weight, then LHS is 0 and RHS is nonnegative.
by_cases A : ∃ i ∈ s, z i = 0 ∧ w i ≠ 0
· rcases A with ⟨i, his, hzi, hwi⟩
rw [prod_eq_zero his]
· exact sum_nonneg fun j hj => mul_nonneg (hw j hj) (hz j hj)
· rw [hzi]
exact zero_rpow hwi
-- If all numbers `z i` with non-zero weight are positive, then we apply Jensen's inequality
-- for `exp` and numbers `log (z i)` with weights `w i`.
· simp only [not_exists, not_and, Ne.def, Classical.not_not] at A
have := convexOn_exp.map_sum_le hw hw' fun i _ => Set.mem_univ <| log (z i)
simp only [exp_sum, (· ∘ ·), smul_eq_mul, mul_comm (w _) (log _)] at this
convert this using 1 <;> [apply prod_congr rfl;apply sum_congr rfl] <;> intro i hi
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· exact rpow_def_of_pos hz _
· cases' eq_or_lt_of_le (hz i hi) with hz hz
· simp [A i hi hz.symm]
· rw [exp_log hz]
#align real.geom_mean_le_arith_mean_weighted Real.geom_mean_le_arith_mean_weighted
/-- AM-GM inequality: the **geometric mean is less than or equal to the arithmetic mean**. --/
theorem geom_mean_le_arith_mean {ι : Type*} (s : Finset ι) (w : ι → ℝ) (z : ι → ℝ)
(hw : ∀ i ∈ s, 0 ≤ w i) (hw' : 0 < ∑ i in s, w i) (hz : ∀ i ∈ s, 0 ≤ z i) :
(∏ i in s, z i ^ w i) ^ (∑ i in s, w i)⁻¹ ≤ (∑ i in s, w i * z i) / (∑ i in s, w i) := by
convert geom_mean_le_arith_mean_weighted s (fun i => (w i) / ∑ i in s, w i) z ?_ ?_ hz using 2
· rw [← finset_prod_rpow _ _ (fun i hi => rpow_nonneg_of_nonneg (hz _ hi) _) _]
refine Finset.prod_congr rfl (fun _ ih => ?_)
rw [div_eq_mul_inv, rpow_mul (hz _ ih)]
· simp_rw [div_eq_mul_inv, mul_assoc, mul_comm, ← mul_assoc, ← Finset.sum_mul, mul_comm]
· exact fun _ hi => div_nonneg (hw _ hi) (le_of_lt hw')
· simp_rw [div_eq_mul_inv, ← Finset.sum_mul]
exact mul_inv_cancel (by linarith)
theorem geom_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = x :=
calc
∏ i in s, z i ^ w i = ∏ i in s, x ^ w i := by
refine' prod_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with h₀ | h₀
· rw [h₀, rpow_zero, rpow_zero]
· rw [hx i hi h₀]
_ = x := by
rw [← rpow_sum_of_nonneg _ hw, hw', rpow_one]
have : (∑ i in s, w i) ≠ 0 := by
rw [hw']
exact one_ne_zero
obtain ⟨i, his, hi⟩ := exists_ne_zero_of_sum_ne_zero this
rw [← hx i his hi]
exact hz i his
#align real.geom_mean_weighted_of_constant Real.geom_mean_weighted_of_constant
theorem arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw' : ∑ i in s, w i = 1)
(hx : ∀ i ∈ s, w i ≠ 0 → z i = x) : ∑ i in s, w i * z i = x :=
calc
∑ i in s, w i * z i = ∑ i in s, w i * x := by
refine' sum_congr rfl fun i hi => _
rcases eq_or_ne (w i) 0 with hwi | hwi
· rw [hwi, zero_mul, zero_mul]
· rw [hx i hi hwi]
_ = x := by rw [← sum_mul, hw', one_mul]
#align real.arith_mean_weighted_of_constant Real.arith_mean_weighted_of_constant
theorem geom_mean_eq_arith_mean_weighted_of_constant (w z : ι → ℝ) (x : ℝ) (hw : ∀ i ∈ s, 0 ≤ w i)
(hw' : ∑ i in s, w i = 1) (hz : ∀ i ∈ s, 0 ≤ z i) (hx : ∀ i ∈ s, w i ≠ 0 → z i = x) :
∏ i in s, z i ^ w i = ∑ i in s, w i * z i := by
rw [geom_mean_weighted_of_constant, arith_mean_weighted_of_constant] <;> assumption
#align real.geom_mean_eq_arith_mean_weighted_of_constant Real.geom_mean_eq_arith_mean_weighted_of_constant
end Real
namespace NNReal
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for `NNReal`-valued functions. -/
theorem geom_mean_le_arith_mean_weighted (w z : ι → ℝ≥0) (hw' : ∑ i in s, w i = 1) :
(∏ i in s, z i ^ (w i : ℝ)) ≤ ∑ i in s, w i * z i :=
mod_cast
Real.geom_mean_le_arith_mean_weighted _ _ _ (fun i _ => (w i).coe_nonneg)
(by assumption_mod_cast) fun i _ => (z i).coe_nonneg
#align nnreal.geom_mean_le_arith_mean_weighted NNReal.geom_mean_le_arith_mean_weighted
/-- The geometric mean is less than or equal to the arithmetic mean, weighted version
for two `NNReal` numbers. -/
theorem geom_mean_le_arith_mean2_weighted (w₁ w₂ p₁ p₂ : ℝ≥0) :
w₁ + w₂ = 1 → p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one] using
geom_mean_le_arith_mean_weighted univ ![w₁, w₂] ![p₁, p₂]
#align nnreal.geom_mean_le_arith_mean2_weighted NNReal.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted (w₁ w₂ w₃ p₁ p₂ p₃ : ℝ≥0) :
w₁ + w₂ + w₃ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃] ![p₁, p₂, p₃]
#align nnreal.geom_mean_le_arith_mean3_weighted NNReal.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted (w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ≥0) :
w₁ + w₂ + w₃ + w₄ = 1 →
p₁ ^ (w₁ : ℝ) * p₂ ^ (w₂ : ℝ) * p₃ ^ (w₃ : ℝ) * p₄ ^ (w₄ : ℝ) ≤
w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ := by
simpa only [Fin.prod_univ_succ, Fin.sum_univ_succ, Finset.prod_empty, Finset.sum_empty,
Finset.univ_eq_empty, Fin.cons_succ, Fin.cons_zero, add_zero, mul_one, ← add_assoc,
mul_assoc] using geom_mean_le_arith_mean_weighted univ ![w₁, w₂, w₃, w₄] ![p₁, p₂, p₃, p₄]
#align nnreal.geom_mean_le_arith_mean4_weighted NNReal.geom_mean_le_arith_mean4_weighted
end NNReal
namespace Real
theorem geom_mean_le_arith_mean2_weighted {w₁ w₂ p₁ p₂ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hw : w₁ + w₂ = 1) : p₁ ^ w₁ * p₂ ^ w₂ ≤ w₁ * p₁ + w₂ * p₂ :=
NNReal.geom_mean_le_arith_mean2_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean2_weighted Real.geom_mean_le_arith_mean2_weighted
theorem geom_mean_le_arith_mean3_weighted {w₁ w₂ w₃ p₁ p₂ p₃ : ℝ} (hw₁ : 0 ≤ w₁) (hw₂ : 0 ≤ w₂)
(hw₃ : 0 ≤ w₃) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃) (hw : w₁ + w₂ + w₃ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ :=
NNReal.geom_mean_le_arith_mean3_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨p₁, hp₁⟩ ⟨p₂, hp₂⟩
⟨p₃, hp₃⟩ <|
NNReal.coe_eq.1 hw
#align real.geom_mean_le_arith_mean3_weighted Real.geom_mean_le_arith_mean3_weighted
theorem geom_mean_le_arith_mean4_weighted {w₁ w₂ w₃ w₄ p₁ p₂ p₃ p₄ : ℝ} (hw₁ : 0 ≤ w₁)
(hw₂ : 0 ≤ w₂) (hw₃ : 0 ≤ w₃) (hw₄ : 0 ≤ w₄) (hp₁ : 0 ≤ p₁) (hp₂ : 0 ≤ p₂) (hp₃ : 0 ≤ p₃)
(hp₄ : 0 ≤ p₄) (hw : w₁ + w₂ + w₃ + w₄ = 1) :
p₁ ^ w₁ * p₂ ^ w₂ * p₃ ^ w₃ * p₄ ^ w₄ ≤ w₁ * p₁ + w₂ * p₂ + w₃ * p₃ + w₄ * p₄ :=
NNReal.geom_mean_le_arith_mean4_weighted ⟨w₁, hw₁⟩ ⟨w₂, hw₂⟩ ⟨w₃, hw₃⟩ ⟨w₄, hw₄⟩ ⟨p₁, hp₁⟩
⟨p₂, hp₂⟩ ⟨p₃, hp₃⟩ ⟨p₄, hp₄⟩ <|
NNReal.coe_eq.1 <| by assumption
#align real.geom_mean_le_arith_mean4_weighted Real.geom_mean_le_arith_mean4_weighted
end Real
end GeomMeanLEArithMean
section Young
/-! ### Young's inequality -/
namespace Real
/-- Young's inequality, a version for nonnegative real numbers. -/
theorem young_inequality_of_nonneg {a b p q : ℝ} (ha : 0 ≤ a) (hb : 0 ≤ b)
(hpq : p.IsConjugateExponent q) : a * b ≤ a ^ p / p + b ^ q / q := by
simpa [← rpow_mul, ha, hb, hpq.ne_zero, hpq.symm.ne_zero, _root_.div_eq_inv_mul] using
geom_mean_le_arith_mean2_weighted hpq.one_div_nonneg hpq.symm.one_div_nonneg
(rpow_nonneg_of_nonneg ha p) (rpow_nonneg_of_nonneg hb q) hpq.inv_add_inv_conj
#align real.young_inequality_of_nonneg Real.young_inequality_of_nonneg
/-- Young's inequality, a version for arbitrary real numbers. -/
theorem young_inequality (a b : ℝ) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ |a| ^ p / p + |b| ^ q / q :=
calc
a * b ≤ |a * b| := le_abs_self (a * b)
_ = |a| * |b| := (abs_mul a b)
_ ≤ |a| ^ p / p + |b| ^ q / q :=
Real.young_inequality_of_nonneg (abs_nonneg a) (abs_nonneg b) hpq
#align real.young_inequality Real.young_inequality
end Real
namespace NNReal
/-- Young's inequality, `ℝ≥0` version. We use `{p q : ℝ≥0}` in order to avoid constructing
witnesses of `0 ≤ p` and `0 ≤ q` for the denominators. -/
theorem young_inequality (a b : ℝ≥0) {p q : ℝ≥0} (hp : 1 < p) (hpq : 1 / p + 1 / q = 1) :
a * b ≤ a ^ (p : ℝ) / p + b ^ (q : ℝ) / q :=
Real.young_inequality_of_nonneg a.coe_nonneg b.coe_nonneg ⟨hp, NNReal.coe_eq.2 hpq⟩
#align nnreal.young_inequality NNReal.young_inequality
/-- Young's inequality, `ℝ≥0` version with real conjugate exponents. -/
theorem young_inequality_real (a b : ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / Real.toNNReal p + b ^ q / Real.toNNReal q := by
nth_rw 1 [← Real.coe_toNNReal p hpq.nonneg]
nth_rw 1 [← Real.coe_toNNReal q hpq.symm.nonneg]
exact young_inequality a b hpq.one_lt_nnreal hpq.inv_add_inv_conj_nnreal
#align nnreal.young_inequality_real NNReal.young_inequality_real
end NNReal
namespace ENNReal
/-- Young's inequality, `ℝ≥0∞` version with real conjugate exponents. -/
theorem young_inequality (a b : ℝ≥0∞) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
a * b ≤ a ^ p / ENNReal.ofReal p + b ^ q / ENNReal.ofReal q := by
by_cases h : a = ⊤ ∨ b = ⊤
· refine' le_trans le_top (le_of_eq _)
repeat rw [div_eq_mul_inv]
cases' h with h h <;> rw [h] <;> simp [h, hpq.pos, hpq.symm.pos]
push_neg at h
-- if a ≠ ⊤ and b ≠ ⊤, use the nnreal version: nnreal.young_inequality_real
rw [← coe_toNNReal h.left, ← coe_toNNReal h.right, ← coe_mul, coe_rpow_of_nonneg _ hpq.nonneg,
coe_rpow_of_nonneg _ hpq.symm.nonneg, ENNReal.ofReal, ENNReal.ofReal, ←
@coe_div (Real.toNNReal p) _ (by simp [hpq.pos]), ←
@coe_div (Real.toNNReal q) _ (by simp [hpq.symm.pos]), ← coe_add, coe_le_coe]
exact NNReal.young_inequality_real a.toNNReal b.toNNReal hpq
#align ennreal.young_inequality ENNReal.young_inequality
end ENNReal
end Young
section HolderMinkowski
/-! ### Hölder's and Minkowski's inequalities -/
namespace NNReal
private theorem inner_le_Lp_mul_Lp_of_norm_le_one (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p ≤ 1) (hg : ∑ i in s, g i ^ q ≤ 1) :
∑ i in s, f i * g i ≤ 1 := by
have hp_ne_zero : Real.toNNReal p ≠ 0 := (zero_lt_one.trans hpq.one_lt_nnreal).ne.symm
have hq_ne_zero : Real.toNNReal q ≠ 0 := (zero_lt_one.trans hpq.symm.one_lt_nnreal).ne.symm
calc
∑ i in s, f i * g i ≤ ∑ i in s, (f i ^ p / Real.toNNReal p + g i ^ q / Real.toNNReal q) :=
Finset.sum_le_sum fun i _ => young_inequality_real (f i) (g i) hpq
_ = (∑ i in s, f i ^ p) / Real.toNNReal p + (∑ i in s, g i ^ q) / Real.toNNReal q := by
rw [sum_add_distrib, sum_div, sum_div]
_ ≤ 1 / Real.toNNReal p + 1 / Real.toNNReal q := by
refine' add_le_add _ _
· rwa [div_le_iff hp_ne_zero, div_mul_cancel _ hp_ne_zero]
· rwa [div_le_iff hq_ne_zero, div_mul_cancel _ hq_ne_zero]
_ = 1 := hpq.inv_add_inv_conj_nnreal
private theorem inner_le_Lp_mul_Lp_of_norm_eq_zero (f g : ι → ℝ≥0) {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : ∑ i in s, f i ^ p = 0) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
simp only [hf, hpq.ne_zero, one_div, sum_eq_zero_iff, zero_rpow, zero_mul,
inv_eq_zero, Ne.def, not_false_iff, le_zero_iff, mul_eq_zero]
intro i his
left
rw [sum_eq_zero_iff] at hf
exact (rpow_eq_zero_iff.mp (hf i his)).left
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. Version for sums over finite sets,
with `ℝ≥0`-valued functions. -/
theorem inner_le_Lp_mul_Lq (f g : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
∑ i in s, f i * g i ≤ (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := by
by_cases hF_zero : ∑ i in s, f i ^ p = 0
· exact inner_le_Lp_mul_Lp_of_norm_eq_zero s f g hpq hF_zero
by_cases hG_zero : ∑ i in s, g i ^ q = 0
· calc
∑ i in s, f i * g i = ∑ i in s, g i * f i := by
congr with i
rw [mul_comm]
_ ≤ (∑ i in s, g i ^ q) ^ (1 / q) * (∑ i in s, f i ^ p) ^ (1 / p) :=
(inner_le_Lp_mul_Lp_of_norm_eq_zero s g f hpq.symm hG_zero)
_ = (∑ i in s, f i ^ p) ^ (1 / p) * (∑ i in s, g i ^ q) ^ (1 / q) := mul_comm _ _
let f' i := f i / (∑ i in s, f i ^ p) ^ (1 / p)
let g' i := g i / (∑ i in s, g i ^ q) ^ (1 / q)
suffices (∑ i in s, f' i * g' i) ≤ 1 by
simp_rw [div_mul_div_comm, ← sum_div] at this
rwa [div_le_iff, one_mul] at this
refine' mul_ne_zero _ _
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hF_zero
· rw [Ne.def, rpow_eq_zero_iff, not_and_or]
exact Or.inl hG_zero
refine' inner_le_Lp_mul_Lp_of_norm_le_one s f' g' hpq (le_of_eq _) (le_of_eq _)
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.ne_zero, rpow_one,
div_self hF_zero]
· simp_rw [div_rpow, ← sum_div, ← rpow_mul, one_div, inv_mul_cancel hpq.symm.ne_zero,
rpow_one, div_self hG_zero]
#align nnreal.inner_le_Lp_mul_Lq NNReal.inner_le_Lp_mul_Lq
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_hasSum`. -/
theorem inner_le_Lp_mul_Lq_tsum {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
(Summable fun i => f i * g i) ∧
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
have H₁ : ∀ s : Finset ι,
∑ i in s, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) := by
intro s
refine' le_trans (inner_le_Lp_mul_Lq s f g hpq) (mul_le_mul _ _ bot_le bot_le)
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hf
· rw [NNReal.rpow_le_rpow_iff (one_div_pos.mpr hpq.symm.pos)]
exact sum_le_tsum _ (fun _ _ => zero_le _) hg
have bdd : BddAbove (Set.range fun s => ∑ i in s, f i * g i) := by
refine' ⟨(∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q), _⟩
rintro a ⟨s, rfl⟩
exact H₁ s
have H₂ : Summable _ := (hasSum_of_isLUB _ (isLUB_ciSup bdd)).summable
exact ⟨H₂, tsum_le_of_sum_le H₂ H₁⟩
#align nnreal.inner_le_Lp_mul_Lq_tsum NNReal.inner_le_Lp_mul_Lq_tsum
theorem summable_mul_of_Lp_Lq {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
Summable fun i => f i * g i :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).1
#align nnreal.summable_mul_of_Lp_Lq NNReal.summable_mul_of_Lp_Lq
theorem inner_le_Lp_mul_Lq_tsum' {f g : ι → ℝ≥0} {p q : ℝ} (hpq : p.IsConjugateExponent q)
(hf : Summable fun i => f i ^ p) (hg : Summable fun i => g i ^ q) :
∑' i, f i * g i ≤ (∑' i, f i ^ p) ^ (1 / p) * (∑' i, g i ^ q) ^ (1 / q) :=
(inner_le_Lp_mul_Lq_tsum hpq hf hg).2
#align nnreal.inner_le_Lp_mul_Lq_tsum' NNReal.inner_le_Lp_mul_Lq_tsum'
/-- Hölder inequality: the scalar product of two functions is bounded by the product of their
`L^p` and `L^q` norms when `p` and `q` are conjugate exponents. A version for `NNReal`-valued
functions. For an alternative version, convenient if the infinite sums are not already expressed as
`p`-th powers, see `inner_le_Lp_mul_Lq_tsum`. -/
theorem inner_le_Lp_mul_Lq_hasSum {f g : ι → ℝ≥0} {A B : ℝ≥0} {p q : ℝ}
(hpq : p.IsConjugateExponent q) (hf : HasSum (fun i => f i ^ p) (A ^ p))
(hg : HasSum (fun i => g i ^ q) (B ^ q)) : ∃ C, C ≤ A * B ∧ HasSum (fun i => f i * g i) C := by
obtain ⟨H₁, H₂⟩ := inner_le_Lp_mul_Lq_tsum hpq hf.summable hg.summable
have hA : A = (∑' i : ι, f i ^ p) ^ (1 / p) := by rw [hf.tsum_eq, rpow_inv_rpow_self hpq.ne_zero]
have hB : B = (∑' i : ι, g i ^ q) ^ (1 / q) := by
rw [hg.tsum_eq, rpow_inv_rpow_self hpq.symm.ne_zero]
refine' ⟨∑' i, f i * g i, _, _⟩
· simpa [hA, hB] using H₂
· simpa only [rpow_self_rpow_inv hpq.ne_zero] using H₁.hasSum
#align nnreal.inner_le_Lp_mul_Lq_has_sum NNReal.inner_le_Lp_mul_Lq_hasSum
/-- For `1 ≤ p`, the `p`-th power of the sum of `f i` is bounded above by a constant times the
sum of the `p`-th powers of `f i`. Version for sums over finite sets, with `ℝ≥0`-valued functions.
-/
theorem rpow_sum_le_const_mul_sum_rpow (f : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, f i) ^ p ≤ (card s : ℝ≥0) ^ (p - 1) * ∑ i in s, f i ^ p := by
cases' eq_or_lt_of_le hp with hp hp
· simp [← hp]
let q : ℝ := p / (p - 1)
have hpq : p.IsConjugateExponent q := by rw [Real.isConjugateExponent_iff hp]
have hp₁ : 1 / p * p = 1 := one_div_mul_cancel hpq.ne_zero
have hq : 1 / q * p = p - 1 := by
rw [← hpq.div_conj_eq_sub_one]
ring
simpa only [NNReal.mul_rpow, ← NNReal.rpow_mul, hp₁, hq, one_mul, one_rpow, rpow_one,
Pi.one_apply, sum_const, Nat.smul_one_eq_coe] using
NNReal.rpow_le_rpow (inner_le_Lp_mul_Lq s 1 f hpq.symm) hpq.nonneg
#align nnreal.rpow_sum_le_const_mul_sum_rpow NNReal.rpow_sum_le_const_mul_sum_rpow
/-- The `L_p` seminorm of a vector `f` is the greatest value of the inner product
`∑ i in s, f i * g i` over functions `g` of `L_q` seminorm less than or equal to one. -/
theorem isGreatest_Lp (f : ι → ℝ≥0) {p q : ℝ} (hpq : p.IsConjugateExponent q) :
IsGreatest ((fun g : ι → ℝ≥0 => ∑ i in s, f i * g i) '' { g | ∑ i in s, g i ^ q ≤ 1 })
((∑ i in s, f i ^ p) ^ (1 / p)) := by
constructor
· use fun i => f i ^ p / f i / (∑ i in s, f i ^ p) ^ (1 / q)
by_cases hf : ∑ i in s, f i ^ p = 0
· simp [hf, hpq.ne_zero, hpq.symm.ne_zero]
· have A : p + q - q ≠ 0 := by simp [hpq.ne_zero]
have B : ∀ y : ℝ≥0, y * y ^ p / y = y ^ p := by
refine' fun y => mul_div_cancel_left_of_imp fun h => _
simp [h, hpq.ne_zero]
simp only [Set.mem_setOf_eq, div_rpow, ← sum_div, ← rpow_mul,
div_mul_cancel _ hpq.symm.ne_zero, rpow_one, div_le_iff hf, one_mul, hpq.mul_eq_add, ←
rpow_sub' _ A, _root_.add_sub_cancel, le_refl, true_and_iff, ← mul_div_assoc, B]
rw [div_eq_iff, ← rpow_add hf, hpq.inv_add_inv_conj, rpow_one]
simpa [hpq.symm.ne_zero] using hf
· rintro _ ⟨g, hg, rfl⟩
apply le_trans (inner_le_Lp_mul_Lq s f g hpq)
simpa only [mul_one] using
mul_le_mul_left' (NNReal.rpow_le_one hg (le_of_lt hpq.symm.one_div_pos)) _
#align nnreal.is_greatest_Lp NNReal.isGreatest_Lp
/-- Minkowski inequality: the `L_p` seminorm of the sum of two vectors is less than or equal
to the sum of the `L_p`-seminorms of the summands. A version for `NNReal`-valued functions. -/
theorem Lp_add_le (f g : ι → ℝ≥0) {p : ℝ} (hp : 1 ≤ p) :
(∑ i in s, (f i + g i) ^ p) ^ (1 / p) ≤
(∑ i in s, f i ^ p) ^ (1 / p) + (∑ i in s, g i ^ p) ^ (1 / p) := by
-- The result is trivial when `p = 1`, so we can assume `1 < p`.
rcases eq_or_lt_of_le hp with (rfl | hp);
· simp [Finset.sum_add_distrib]
have hpq := Real.isConjugateExponent_conjugateExponent hp
have := isGreatest_Lp s (f + g) hpq
simp only [Pi.add_apply, add_mul, sum_add_distrib] at this
rcases this.1 with ⟨φ, hφ, H⟩
rw [← H]
exact
add_le_add ((isGreatest_Lp s f hpq).2 ⟨φ, hφ, rfl⟩) ((isGreatest_Lp s g hpq).2 ⟨φ, hφ, rfl⟩)
#align nnreal.Lp_add_le NNReal.Lp_add_le
/-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) := by
have pos : 0 < p := lt_of_lt_of_le zero_lt_one hp
have H₁ : ∀ s : Finset ι,
(∑ i in s, (f i + g i) ^ p) ≤
((∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p)) ^ p := by
intro s
rw [← NNReal.rpow_one_div_le_iff pos]
| refine' le_trans (Lp_add_le s f g hp) (add_le_add _ _) | /-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) := by
have pos : 0 < p := lt_of_lt_of_le zero_lt_one hp
have H₁ : ∀ s : Finset ι,
(∑ i in s, (f i + g i) ^ p) ≤
((∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p)) ^ p := by
intro s
rw [← NNReal.rpow_one_div_le_iff pos]
| Mathlib.Analysis.MeanInequalities.499_0.4hD1oATDjTWuML9 | /-- Minkowski inequality: the `L_p` seminorm of the infinite sum of two vectors is less than or
equal to the infinite sum of the `L_p`-seminorms of the summands, if these infinite sums both
exist. A version for `NNReal`-valued functions. For an alternative version, convenient if the
infinite sums are already expressed as `p`-th powers, see `Lp_add_le_hasSum_of_nonneg`. -/
theorem Lp_add_le_tsum {f g : ι → ℝ≥0} {p : ℝ} (hp : 1 ≤ p) (hf : Summable fun i => f i ^ p)
(hg : Summable fun i => g i ^ p) :
(Summable fun i => (f i + g i) ^ p) ∧
(∑' i, (f i + g i) ^ p) ^ (1 / p) ≤
(∑' i, f i ^ p) ^ (1 / p) + (∑' i, g i ^ p) ^ (1 / p) | Mathlib_Analysis_MeanInequalities |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.