This paper studies the minimum mean squared error (MMSE) of estimating
$\mathbf{X} \in \mathbb{R}^d$ from the noisy observation $\mathbf{Y} \in
\mathbb{R}^k$, under the assumption that the noise (i.e.,
$\mathbf{Y}|\mathbf{X}$) is a member of the exponential family. The paper
provides a new lower bound on the MMSE. Towards this end, an alternative
representation of the MMSE is first presented, which is argued to be useful in
deriving closed-form expressions for the MMSE. This new representation is then
used together with the Poincar\'e inequality to provide a new lower bound on
the MMSE. Unlike, for example, the Cram\'{e}r-Rao bound, the new bound holds
for all possible distributions on the input $\mathbf{X}$. Moreover, the lower
bound is shown to be tight in the high-noise regime for the Gaussian noise
setting under the assumption that $\mathbf{X}$ is sub-Gaussian. Finally,
several numerical examples are shown which demonstrate that the bound performs
well in all noise regimes.
† New Jersey Institute of Technology, Newark, NJ 07102, USA, Email: {ihz3,alex.dytso}@njit.edu
ニュージャージー工科大学, Newark, NJ 07102, USA, Email: {ihz3, Alex.dytso}@njit.edu
0.77
⋆ University of Minnesota, Minneapolis, MN 55404, USA, Email: mcardone@umn.edu
ミネソタ大学ミネアポリス校, MN 55404, USA, Email: mcardone@umn.edu
0.72
Ian Zieder†, Alex Dytso†, Martina Cardone⋆
イアン・ジーデル、アレックス・ディッツォ、マルティナ・カルドーネ
0.33
Abstract—This paper studies the minimum mean squared error (MMSE) of estimating X ∈ Rd from the noisy observation Y ∈ Rk, under the assumption that the noise (i.e., Y|X) is a member of the exponential family.
抽象 — 本論文は、雑音(すなわち Y|X)が指数族のメンバーであるという仮定の下で、雑音観測 Y ∈ Rk から X ∈ Rd を推定する最小平均二乗誤差(MMSE)を研究する。
0.82
The paper provides a new lower bound on the MMSE.
論文はMMSEの新たな下限を提供する。
0.62
Towards this end, an alternative representation of the MMSE is first presented, which is argued to be useful in deriving closed-form expressions for the MMSE.
This new representation is then used together with the Poincar´e inequality to provide a new lower bound on the MMSE.
この新しい表現はポアンカルの不等式と共にmmseに新しい下界を与えるために使われる。
0.62
Unlike, for example, the Cram´er-Rao bound, the new bound holds for all possible distributions on the input X. Moreover, the lower bound is shown to be tight in the high-noise regime for the Gaussian noise setting under the assumption that X is subGaussian.
例えば、クレーア・ラオ境界とは異なり、入力 x 上の任意の可能な分布に対する新しいバウンドは、x が準ガウジアンであると仮定したガウス雑音設定の高雑音環境において、下限が密であることを示している。
0.68
Finally, several numerical examples are shown which demonstrate that the bound performs well in all noise regimes.
最後に、バウンドがすべてのノイズレジームでうまく機能することを示すいくつかの数値例を示す。
0.68
I. INTRODUCTION
I. イントロダクション
0.64
The minimum mean squared error (MMSE) is an essential and ubiquitous fidelity criterion in statistical signal processing.
最小平均二乗誤差(MMSE)は統計信号処理において必須かつユビキタスな忠実度基準である。
0.78
However, the MMSE is often difficult to compute in closedform, and we often need to rely on bounds.
しかし、mmse は閉形式で計算することがしばしば困難であり、しばしば境界に依存する必要がある。
0.74
In terms of bounds, the attention typically falls on lower bounds as deriving a tight lower bound can often be a difficult task.
境界の観点では、通常は下限に注意が向けられるが、下限の厳密な導出は難しい作業であることが多い。
0.67
In this work, we derive a novel lower bound on the MMSE
本研究では,MMSEに基づく新しい下界を導出する。
0.64
inequality, which establishes the following variational representation of the MMSE,
以下に示すMMSEの変動表現を確立する不等式
0.59
mmse(X|Y) = sup
mmse(X|Y) = sup
0.46
ψ∈C kE[ψ(X, Y)XT]k E[kψ(X, Y)k2]
ψψc kE[t(X, Y)XT]kE[t](X, Y)k2]
0.29
, where C = {ψ :X × Y → X : E[ψ(X, Y)|Y = y] = 0, y ∈ Y,
, どこに C = {n :X × Y → X : E[n(X, Y)|Y = y] = 0, y ∈ Y,
0.51
E[kψ(X, Y)k2] < ∞}.
E[k](X, Y)k2] < ∞} である。
0.91
The aforementioned lower bounds are then attained by a clever choice of the function ψ that results in a computationally feasible bound.
上記の下界は関数 ψ の巧妙な選択によって達成され、計算可能な境界となる。
0.60
One of the drawbacks of this family of bounds is that choosing the right ψ can be challenging.
この境界の族における欠点の1つは、正しい ψ を選択することが難しいことである。
0.60
In particular, to the best of our knowledge, all of the existing bounds require that the random vector X has a pdf; as such, these bounds do not, for example, hold for discrete or mixed random vectors.
特に、我々の知る限りでは、既存のすべての境界は確率ベクトル X が pdf を持つことを要求する。 訳抜け防止モード: 特に、我々の知る限りでは、既存のすべての境界は、ランダムベクトル X が pdf を持つ必要がある。 したがって、これらの境界は、例えば、離散または混合ランダムベクトルを保たない。
0.69
The second family of lower bounds is known as Ziv-Zakai and it was originally proposed in [6] and later improved in [7], [8], [9].
Ziv-Zakai bounds rely on connecting estimation to binary hypothesis testing.
ziv-zakai境界は、推定とバイナリ仮説テストの結合に依存している。
0.45
A simple form of the Ziv-Zakai bound can be stated as follows,
Ziv-Zakai 境界の単純形式を次のように述べることができる。
0.67
of estimating X ∈ Rd from the noisy observation Y ∈ Rk.
雑音観測 Y ∈ Rk から X ∈ Rd を推定する。
0.74
Towards this end, we present and study an alternative representation of the MMSE mmse(X|Y).
この目的に向けて、MMSE mmse(X|Y)の代替表現を提示し、研究する。
0.73
This new representation provides a new line of attack for direct computation of the MMSE and, together with the Poincar´e inequality, allows us to derive a new lower bound on the MMSE.
The focus is on the exponential family, which is described next.
次に説明する指数族に焦点が当てられている。
0.59
The class of
class (複数形 classs)
0.44
probability models P = (cid:8)PY|X=x, x ∈ X ⊆ Rd(cid:9) supported on y ∈ Y ⊆ Rk is an exponential family if the probability
確率モデル P = (cid:8)PY|X=x, x ∈ X > Rd(cid:9) が y ∈ Y > Rk 上で成り立つ確率モデルは、確率が指数群であれば指数族である。 訳抜け防止モード: 確率モデル P = ( cid:8)PY|X = x, x ∈ X > Rd(cid:9 ) Rk が指数族である確率
0.73
density function (pdf) of it can be written as
その密度関数(pdf)は、書くことができる
0.84
fY|X(y|x) = h
fY|X(y|x) = h
0.37
(y)ehx,T (y)i−φ
(y)ehx,T (y)i−φ
0.42
(x), (1) where T : Y → Rd is the sufficient statistic function; φ : X → R is the log-partition function; h : Y → [0,∞) is the base measure; and h·,·i denotes the inner product.
(x) (1) ここで t : y → rd は十分統計関数、 φ : x → r は対数分解関数、h : y → [0,∞) は基底測度、h·,·i は内積を表す。 訳抜け防止モード: (x) (1) ここで t : y → rd は十分統計関数である ; φ : x → r は log - partition function ; h : y → [ 0,∞ ) は基底測度 ; そして h·,·i は内積を表す。
0.49
A. Related Work The first
A.関連業務 第一に
0.73
Generally, there are three approaches for finding lower bounds on the MMSE, which result in three different families.
一般的には mmse上の下限を見つけるには3つのアプローチがあり、3つの異なる家族を成す。
0.63
family is known as Weiss–Weinstein family [1], and it includes important bounds such as the Bayesian Cram´er-Rao bound [2] (also known as the Van Trees bound), the Bobrovsky–Zakai bound [3], the Barankin bound [4], and the Bobrovsky-Mayer-Wolf -Zakai bound [5].
The Weiss–Weinstein family relies on the Cauchy-Schwarz
ヴァイス=ワインシュタイン家はコーシー=シュワルツに依存している
0.27
mmse(X|Y) 2Z ∞ 0 Z ∞ ≥
mmse(X|Y) 2Z ∞ 0 Z ∞ ≥
0.46
1 −∞ (fX (x) + fX (x + h)) Pe(x, x + h)hdxdh,
1 −∞ (fX (x) + fX (x + h)) Pe(x, x + h)hdxdh,
0.41
where Pe(x, x + h) is the optimal probability of error for the following binary hypothesis problem,
ここで Pe(x, x + h) は次の二項仮説問題に対する誤差の最適確率である。
0.85
where H0 :Y ∼ PY|X (y|x), H1 :Y ∼ PY|X (y|x + h),
どこに h0 :y はpy|x(y|x)、h1 :y はpy|x(y|x + h)である。
0.60
Pr(H0) = fX(x)
Pr(H0) = fX(x)
0.67
fX (x) + fX (x + h)
fX (x) + fX (x + h)
0.42
, Pr(H1) = 1 − Pr(H0),
, Pr(H1) = 1 − Pr(H0)
0.44
with fX being the pdf of X. While this family of lower bounds is typically very tight, it suffers from several drawbacks.
fx は x の pdf である。この下限の族は一般に非常に密接であるが、いくつかの欠点がある。
0.73
First, it can be difficult to compute in closed-form.
まず、クローズドフォームで計算するのは困難である。
0.68
Second, while there are vector generalizations of this bound [10], typically, these generalizations contain another layer of optimization, which can make the computation difficult.
Third, these bounds assume that X has a density and cannot be used to study the MMSE of discrete or mixed random variables.
第三に、これらの境界は x が密度を持ち、離散確率変数や混合確率変数の mmse を研究できないと仮定する。
0.76
The third family of lower bounds uses a variational approach and it works by minimizing the MSE subject to a constraint on a suitably chosen divergence measure, for example, the
Similar to the previous bounds, also this family only holds if X has a density and hence, it is not suitable for studying the MMSE of discrete or mixed random variables.
以前の境界と同様に、この族は X が密度を持つ場合にのみ成り立つので、離散確率変数や混合確率変数の MMSE の研究には適さない。
0.71
The key ingredient in the proof of our new lower bound on the MMSE is the Poincar´e inequality [12].
The Poincar´e inequality has found a number of applications in information theory and signal processing and the interested reader is referred to [13], [14], [15], [16], [17], [18], [19], [20] and references therein.
B. Contributions and Outline The contributions and the outline of the paper are as follows.
b. 貢献と概要 論文の貢献と概要は以下のとおりである。
0.68
The rest of this section presents relevant notation.
本項の残りは関連する表記である。
0.66
Section II presents the new lower bound on the MMSE, and it shows its tightness in the high-noise regime for the Gaussian noise setting under the assumption that X is sub-Gaussian.
Finally, Section III is dedicated to showing the proof of a new representation of the MMSE, that was used in the proof of the lower bound.
最後に、第III節は、下界の証明に使われたMMSEの新しい表現の証明を示すことに特化している。
0.63
Section III also argues about the computation advantages of this new representation via an example.
第3節は、この新しい表現の計算上の利点を例を通して論じている。
0.59
C. Notation Deterministic scalar quantities are denoted by lowercase letters, scalar random variables are denoted by uppercase letters, vectors are denoted by bold lowercase letters, random vectors by bold uppercase letters, and matrices by bold uppercase sans serif letters.
The smallest eigenvalue of a matrix A is denoted by λmin(A) and the smallest singular value is denoted by σmin(A); Tr(A) is the trace of A; the left inverse of A with full column rank is defined and denoted by A+ = (ATA)−1AT.
行列 A の最小固有値は λmin(A) で表され、最小特異値は σmin(A) で表される; Tr(A) は A のトレースであり、全列ランクの A の左逆数は A+ = (ATA)−1AT で表される。
0.80
Ik is the identity matrix of dimension k, and 0k is the column vector of dimension k of all zeros.
Ik は次元 k の恒等行列であり、0k はすべての零点の次元 k の列ベクトルである。
0.83
The gradient of a function f : Rn → R is denoted by ∇xf (x) =h ∂f (x)
函数 f : Rn → R の勾配は、xf (x) =h ∂f (x) で表される。
0.90
∂xn iT The Jacobian matrix of a function f : Rn → Rm is denoted by Jxf (x) ∈ Rn×m and defined as Jxf (x) =(cid:2)∇xf1(x) ∇xf2(x) The log function is the natural logarithm; h·,·i denotes the inner product; for a function g, we let L{g}(·) denote the Laplace transform of g.
∂xn である。 函数 f : Rn → Rm のジャコビアン行列は Jxf (x) ∈ Rn×m で表され、Jxf (x) = (cid:2) ^xf1(x) ^xf2(x) と定義される: ログ函数は自然対数である; h·,·i は内積を表し、函数 g に対して L{g}(·) は g のラプラス変換を表す。
0.65
. . . ∇xfm(x)(cid:3) .
. . . は (x)(cid:3)。
0.55
∈ Rn. ∂f (x)
∈ rn である。 ∂f(x)
0.54
∂f (x) . . .
∂f(x) . . .
0.46
∂x1 ∂x2 (2)
∂x1 ∂x2 (2)
0.34
(3) II. A NEW LOWER BOUND ON THE MMSE
(3) II。 マンスに作用する新しい低層地盤
0.54
In this section, we derive a novel lower bound on the MMSE.
本項では,MMSEに基づく新しい下限を導出する。
0.54
Towards this end, we first show a new representation of the MMSE, the base of which is the following identity,
この目的に向けて、まず最初にMMSEの新たな表現を示し、その基礎は次のアイデンティティである。
0.70
JyT(y)E[X|Y = y] = ∇y log
JyT(y)E[X|Y = y] = y log
0.47
fY(y) h(y)
fY(y) h(y)
0.42
, y ∈ Y,
, y ∈ y である。
0.74
(4) where it is assumed that PY|X belongs to the exponential family in (1) with sufficient statistics T and base measure h.
(4) ここで PY|X は (1) において十分統計量 T と基底測度 h を持つ指数族に属すると仮定される。
0.60
The identity in (4) is well-known in both statistic and information theory and it often goes under the name of Tweedie’s formula [22], [23].
(4) の同一性は統計学と情報理論の両方でよく知られており、しばしば tweedie's formula [22], [23] の名で表される。
0.83
The identity in (4) can be restated using the information density, which for the distribution PX,Y supported on X × Y is defined as
X × Y でサポートされた分布 PX,Y については、情報密度を用いて (4) のアイデンティティを再定義することができる。
0.89
ιPX,Y (x; y) = log
ιPX,Y (x; y) = log
0.46
(x, y), x ∈ X , y ∈ Y,
(x, y), x ∈ X , y ∈ Y,
0.40
(5) dPX,Y d(PX · PY)
(5) dPX,Y d(PX · PY)
0.43
dPX,Y d(PX·PY) (x, y) =
dPX,Y d(PX·PY) (x, y) =
0.49
(x, y) is the Radonwhere Nikodym derivative.
(x, y) は Nikodym 誘導体である。
0.69
Note that if PY|X=x is not absolutely continuous with respect to PY we let ιPX,Y (x; y) = ∞.
PY|X=x が PY に対して絶対連続でないなら、ιPX,Y (x; y) = ∞ とする。
0.87
In order to derive our new lower bound on the MMSE, we will leverage the alternative expression for the MMSE given in the next theorem, the proof of which is provided in Section III.
Theorem 1, unlike the classical representation, requires only the knowledge of fY|X (which is known) and of the marginal pdf fY.
古典的表現とは異なり、定理 1 は fY|X (既知の) の知識と限界 pdf fY の知識のみを必要とする。
0.77
Thus, such an alternative representation is potentially easier to handle; in Section III, we will show an example of how such an alternative representation can lead to closed-form expressions for the MMSE.
The remaining of this section is dedicated to the derivation of a novel lower bound on the MMSE and its analysis.
本節の残りは、MMSEに基づく新規な下界の導出とその解析に充てられている。
0.62
In particular, the main tool for obtaining our bound is the Poincar´e inequality, which is formally discussed next.
特に、我々の境界を得る主なツールはポアンカルの不等式であり、これは次に正式に議論される。
0.58
A. Poincar´e Inequality Consider a class of functions A. We say that a probability distribution PU satisfies a Poincar´e inequality with respect to A with a constant κ ≥ 0 if [12]
ポアンカルの不等式 確率分布 pu が、定数 κ ≥ 0 の a に対して poincar ′e の不等式を満たすことを [12] とする。
0.54
Var(f (U)) ≤
Var(f (U)) ≤
0.43
1 κ E(cid:2)||∇f (U)||2(cid:3) , ∀f ∈ A.
1 κ E(cid:2)||\f (U)||2(cid:3) , sf ∈ A である。
0.58
(6) If κ = 0, we treat the right-hand side of (6) as infinity.
(6) κ = 0 であれば、(6) の右辺を無限大として扱う。
0.60
We are here interested in the conditional version of the Poincar´e inequality, i.e., for a class of functions A we say that a conditional probability PY|X=x (for a fixed x ∈ X ) satisfies a Poincar´e inequality in (6) with respect to A with a constant κ(x) ≥ 0 if E(cid:2)||∇f (Y)||2|X = x(cid:3) , ∀f ∈ A. Var(f (Y)|X = x) ≤ Since x ∈ X can be treated as a parameter of the distribution, the conditional and unconditional versions hold under the same conditions.
ここではポアンカーの不等式、すなわち函数のクラス A に対して条件確率 PY|X=x (固定 x ∈ X ) が定数 κ(x) ≥ 0 if E(cid:2)||\f(Y)||2|X = x(cid:3) , shf ∈ A. Var(f(Y)|X = x) ≤ x ∈ X は分布のパラメータとして扱うことができるので、条件確率 PY|X=x (固定 x ∈ X ) は A に対して定数 κ(x) ≥ 0 を満たす。
0.76
There exist several sufficient conditions on A and
a にはいくつかの十分な条件が存在する
0.64
κ(x) 1
κ(x) 1
0.43
英語(論文から抽出)
日本語訳
スコア
PY|X, which guarantee that a Poincar´e inequality holds, and which identify the constant κ(x).
PY|X はポアンカーの不等式が成り立つことを保証し、定数 κ(x) を識別する。
0.71
We next list a few of these.
次にそのいくつかを挙げる。
0.54
• Convex Poincar´e [26].
・Convex Poincar ́e [26]
0.38
Let PY|X be a product distribution and A be a set of functions such that f ∈ A is f : [0, 1]k → R, separately convex and the partial derivatives of which exist.
PY|X を積分布とし、A を f ∈ A が f : [0, 1]k → R となるような函数の集合とする。 訳抜け防止モード: py|x を積分布とし、a を f ∈ a が f : [ 0, となるような函数の集合とする。 1]k → r, 別個の凸と部分微分が存在する。
0.78
Then, κC(x) = 1.
すると κC(x) = 1 となる。
0.81
• Bakry- ´Emery condition [27]: Let A be a class of contin-
• Bakry- ́Emery condition [27]: A をContin のクラスとする
0.77
uously differentiable functions.
明確に区別できる機能です
0.51
Then, κBE(x)
そしたら κBE(x)
0.55
= max(cid:26)κ :∇2
= max(cid:26)κ :=2
0.43
y log(cid:18)
y log(cid:18)
0.44
1 fY|X(y|x)(cid:19)(cid:23) κIk,∀y∈Y(cid:27) .
1 fy|x(y|x)(cid:19)(cid:23) κik(cid:27) である。
0.51
(7) If the set in (7) is empty, then we set κBE(x)
(7) 7) の集合が空であれば、κBE(x) をセットする。
0.59
= 0. We note that the Bakry- ´Emery condition in (7) simply requires that the distribution is strongly log-concave.
As an example, the Bakry- ´Emery constant for the exponential family is given by the next proposition.
例えば、指数関数族に対するbakry-'emery定数は次の命題によって与えられる。
0.57
Proposition 1. Assume that PY|X has a pdf of the form in (1).
提案 1. PY|X が (1) の形の pdf を持つと仮定する。
0.54
Then, for x ∈ X , we have κBE(x) = max{0, ˜κBE(x)} , y log(cid:18) 1 λmin(cid:18)∇2 ˜κBE(x) = min y∈Y
すると、x ∈ x に対して、κbe(x) = max{0, \κbe(x)} と y log(cid:18) 1 λmin(cid:18) が成立する。
0.83
yhx, T(y)i(cid:19) ,
yhx, T(y)i(cid:19) ,
0.50
h(y)(cid:19)−∇2
h(y)(cid:19)−2
0.43
y denotes the Hessian.
yはヘシアンを表す。
0.65
Proof: We have where ∇2 y log(cid:18) ∇2 = −∇2 (cid:23) λmin(cid:18)∇2
証明: 2 y log(cid:18) が −2 (cid:23) λmin(cid:18) である場合
0.31
fY|X(y|x)(cid:19) y(cid:0)log (h(y)) − ∇2 y log(cid:18) 1
fY|X(y|x)(cid:19) y(cid:0)log (h(y)) − y log(cid:18) 1
0.43
1 y(hx, T
1 y(hx, T)
0.45
(y)i) − φ (x)(cid:1) yhx, T
(y)i) − φ (x)(cid:1)yhx,T
0.45
(y)i(cid:19) Ik,
(y)i(cid:19)Ik,
0.48
h(y)(cid:19) − ∇2
h(y)(cid:19) − は 2 である。
0.64
which concludes the proof of Proposition 1.
これは命題1の証明を結論付ける。
0.54
• Laplace distribution [28]: Let PY |X=x have a Laplace pdf (i.e., fY |X (y|x) = 1 2 e−|y−x|) and A be a set of all functions f : R → R that are continuously differentiable and limx±∞ e−|x|f (x) = 0.
• laplace distribution [28]: py |x=x をラプラス pdf(すなわち fy |x (y|x) = 1 2 e−|y−x|) とし、連続微分可能かつ limx±∞ e−|x|f (x) = 0 であるすべての函数 f : r → r の集合とする。
0.82
Then, κLap(x) = 1 4 .
そして κLap(x) = 1 4 となる。
0.89
B. A New Lower Bound on the MMSE
B.MMSEにおける新しい下界
0.63
We here leverage the result in Theorem 1 to derive a new lower bound on the MMSE for the exponential family (i.e., we assume that PY|X has a pdf of the form in (1)).
ここでは、定理 1 の結果を利用して指数族に対する MMSE 上の新しい下界を導出する(つまり、PY|X は形式 (1) の pdf を持つと仮定する)。
0.76
Our new lower bound on the MMSE is given by the next theorem.
MMSEの新たな下界は次の定理によって与えられる。
0.67
Theorem 2. Assume that the following three conditions hold: 1) For all x ∈ X the distribution PY|X=x has a pdf of the form in (1) and it satisfies a Poincar´e inequality with respect to (A, κ
定理2。 1) すべての x ∈ x に対して、分布 py|x=x は (1) の形式の pdf を持ち、(a, κ) に関して poincar ′e の不等式を満たすと仮定する。
0.70
(x)); 2) y 7→ ιPXY (x;
(x)); 2) y 7→ ιPXY (x;
0.43
y) ∈ A for every x such that κ
y) κ が成り立つすべての x に対する ∈ a
0.92
(x) > 0; 3) There exists a ρ ≥ 0 such that for all y ∈ Y, σmin(cid:16)(JyT
(x) > 0; 3) すべての y ∈ Y, σmin(cid:16)(JyT) に対して ρ ≥ 0 が存在する。
0.94
(y))+(cid:17) ≥ ρ.
(y)+(cid:17) ≥ ρ である。
0.85
mmse(X|Y) ≥ ρ2E [κ(X)Var(ιPXY (X; Y)|X)] .
mmse(x|y) ≥ ρ2e [κ(x)var(ιpxy (x; y)|x)] である。
0.69
Then, Proof: We have mmse(X|Y) = E(cid:2)||X − E[X|Y]||2(cid:3)
(b) using condition 3) in Theorem 2 and the inequality kAxk ≥ σmin(A)kxk; and
(b)定理2における条件3と不等式kaxk ≥ σmin(a)kxk
0.72
(c) using a Poincar´e inequality and conditions
(c)ポアンカーの不平等と条件の使用
0.73
1) and 2) in Theorem 2.
1)と 2) 定理2。
0.67
This concludes the proof of Theorem 2.
これは定理2の証明を結論付ける。
0.73
Remark 1. Theorem 2 holds provided that three conditions are satisfied.
備考1。 定理2では、3つの条件が満たされる。
0.57
Conditions 1) and 2) are required for the application of a Poincar´e inequality in the proof of the bound.
条件 1)と 2) 境界の証明において poincar ′e の不等式の適用は必須である。
0.75
Specifically, in Section II-A, we have listed a number of sufficient conditions for
具体的には、第ii-a節において、我々は多くの十分な条件を列挙した。
0.48
1) to hold. Condition 2) requires that the information density y 7→ ιPXY (x; y) belongs to some regular enough family of functions A. Interestingly, such conditions are not difficult to find.
1) 保持。 条件 2) 情報密度 y 7 → ιPXY (x; y) は、ある正規関数の族 A に属することを要求する。 訳抜け防止モード: 1) 保持。 条件 2 ) は情報密度 y 7 → ιPXY ( x ; y ) が任意の正規関数族 A に属することを要求する。 このような条件は 見つけるのが難しくありません
0.60
Moreover, often these conditions only depend on PY|X and are independent of PX.
さらに、これらの条件はしばしば PY|X にのみ依存し、PX とは独立である。
0.68
For example, the information density for the exponential family in (1) is known to be infinitely differentiable for all distributions on PX [25].
Finally, condition 3) imposes a requirement on the sufficient statistics T(y).
最後に、条件 3 は十分な統計量 T(y) に要件を課す。
0.81
This condition, for example, holds when T(y) is a linear function (e g , Gaussian, Wishart).
この条件は、例えば、T(y) が線型函数(例えば、ガウス、ウィッシュアート)であるときに成立する。
0.74
C. Tightness in the High-Noise Regime
c. 高音域におけるタイトネス
0.54
We here show an example of PY|X for which our lower bound in Theorem 2 is tight in the high-noise regime.
ここでは、定理 2 における下限が高雑音状態においてタイトであるような py|x の例を示す。
0.54
Towards this end, we consider a scenario where
この目的に向けて、我々はシナリオを考える。
0.66
Y = X + N,
y = x + n である。
0.71
where X and N are independent and N ∼ N (0k, σ2
X と N は独立であり、N は N (0k, σ2) である。
0.73
this case, the lower bound in Theorem 2 reduces to
この場合、Theorem 2の下位境界は減少する
0.74
N (8) Ik).
N (8) IK)。
0.52
In mmse(X|Y) ≥ σ2
院 mmse(X|Y) ≥ σ2
0.42
N E [Var(ι(X; Y)|X)] .
N e [var(ι(x; y)|x)] である。
0.56
(9) It is noted that (9) is a new representation of the MMSE.
(9) 9)はMMSEの新しい表現である。
0.44
Conditions 1)-3) are verified as follows.
条件 1)-3)は以下のとおりである。
0.82
First, we note that for the model in (8), we have that T(y) = y/σ2 N , which implies ρ = σ2 N in condition 3).
まず、(8) のモデルに対して、t(y) = y/σ2 n であり、これは条件 3) において ρ = σ2 n を意味する。
0.77
To verify conditions 1) and 2), we use the Bakry- ´Emery condition presented in Section II-A, which first requires that y 7→ ιPXY (x; y) is continuously differentiable for every x, which is a well-known fact, see for example
条件 1) と 2) を検証するために、第2-A節で示されるBakry- ́Emery条件を使い、まずは y 7 → ιPXY (x; y) がすべての x に対して連続的に微分可能であることを要求する。
0.75
[24]. Second, since h
[24]. 第二にhは
0.54
(y) = N , we can find the Bakry- ´Emery constant by applying Proposition 1, namely κ
(y) = N , Bakry- ́Emery 定数は、命題 1、すなわち κ を適用することによって見つけることができる。
0.67
(x) = κBE (x) = 1/σ2 N .
(x) =κBE (x) = 1/σ2 n である。
0.56
The quantity E [Var(ι(X; Y)|X)] has appeared in the past in [29], where it was termed as conditional information variance.
e [var(ι(x; y)|x)] は [29] において以前に現れており、条件付き情報分散 (conditional information variance) と呼ばれた。
0.78
N )k/2 e 1 (2πσ2
N)k/2 e 1 (2πσ2
0.38
− kyk2 2σ2
-kyk2 2σ2
0.55
英語(論文から抽出)
日本語訳
スコア
We now argue that the bound in (9) is tight in the highnoise regime.
現在、(9)のバウンドはハイノワーズ体制において厳密であると主張する。
0.68
To do this, we recall the following high-noise behavior of the MMSE in the Gaussian noise setting [30]
これを実現するために、ガウス雑音設定におけるMMSEの次の高雑音挙動を思い出す[30]。
0.74
We now leverage the following lemma, the proof of which is provided in Appendix A.
現在、以下の補題を利用しており、その証明は Appendix A で提供されている。
0.61
Lemma 1. Assume that X is sub-Gaussian.
レマ1号。 X がガウス以下であると仮定する。
0.50
Then, lim σN →∞
そしたら リム σn →∞ である。
0.48
mmse(X|Y) = Var (X) .
mmse(X|Y) = Var (X) である。
0.80
At this point, it is interesting to point out that for the Gaussian noise setting, the Cram´er-Rao bound is given by [2]
100 σ2 N Fig. 1: X ∼ N (06, ΣX) where ΣX is given in (18).
100 σ2 N 図1は、 ΣX が 18 で与えられるような ΣX (06, ΣX) である。
0.55
D. Numerical Evaluations The result in Theorem 3 shows that our lower bound in Theorem 2 is tight in the high-noise regime for a large family of prior distributions on X. However, we suspect that such a tightness result holds more generally.
D.数値評価 Theorem 3 の結果は、我々の定理 2 の下位境界が X 上の先行分布の大きい族に対する高雑音状態にあることを示しているが、そのような厳密性の結果はより一般的に成り立つと推測する。
0.82
To support this fact, we here provide numerical evaluations for three ‘toy’, yet practically relevant, scenarios.
この事実を裏付けるために、我々は3つの「トイ」シナリオについて数値的評価を提供する。
0.71
1) Gaussian Input. We assume that X ∼ N (0k, ΣX), i.e., X is a Gaussian random vector.
(16) and the Cram´er-Rao lower bound evaluates to [2]
(16) Cram ́er-Rao の下限は [2] と評価される
0.53
mmse(X|Y) ≥
mmse(X|Y) ≥
0.44
k2 k σ2 N
k2 k σ2 n である。
0.45
+ Tr[Σ−1 X ]
+Tr[Σ−1 X ]
0.44
. (17) In Fig 1, we plot the MMSE in (16) (solid line), the Cram´er-Rao bound in (17) (dotted line), and our bound on the MMSE in (9) (dashed line) versus different values of σ2 N for k = 6 and a randomly generated ΣX, given by
Finally, we point out that for the scenarios of a BPSK input (Fig. 2) and a sparse input (Fig. 3), i.e., where X has a discrete or a mixed distribution, commonly used lower bounds (e g , Cram´er-Rao) do not hold, whereas ours does.
Hence, it would also be interesting to characterize the behavior of our lower bound in the low-noise regime.
したがって、低雑音状態における我々の下界の挙動を特徴付けることも興味深い。
0.66
However, in the low-noise regime, the MMSE has an intricate behavior, for example, it depends on whether the distribution of X is discrete or continuous [37]; we leave this direction for future work.
しかし、低雑音環境では、例えば、mmse は複雑な振る舞いを持ち、それは x の分布が離散的であるか連続的であるか [37] に依存する。 訳抜け防止モード: しかし、低騒音では。 MMSEは複雑な振る舞いをしています 例えば、それは X の分布は離散的か連続的である[37 ] 私たちはこの方向を将来の仕事に置き去りにします。
0.68
III. PROOF OF THEOREM 1 AND ITS APPLICABILITY
III。 理論1のプローフとその適用性
0.68
In this section, we prove Theorem 1, which provides a new representation of the MMSE.
本稿では,MMSE の新たな表現を提供する Theorem 1 を証明する。
0.72
Towards this end, we leverage the following proposition, which provides an expression for the gradient of the information density.
この目的に向けて,情報密度の勾配を表す式として,以下の提案を利用する。
0.70
Proposition 2. For x ∈ X , y ∈ Y, we have that
第2話。 x ∈ X , y ∈ Y に対して、それはある。
0.64
∇yιPXY (x; y) = JyT(y)(x − E[X|Y = y]).
yιPXY (x; y) = JyT(y)(x − E[X|Y = y])。
0.45
英語(論文から抽出)
日本語訳
スコア
err
翻訳エラー
0.00
英語(論文から抽出)
日本語訳
スコア
By substituting the above inside (28), we obtain
上記(28)を置換することにより、我々は、
0.72
mmse(X|Y ) =
mmse(X|Y ) =
0.44
α(α + 1) 2(cid:1) .
α(α + 1) 2(cid:1)であった。
0.47
β2(cid:0)α + 3
β2(cid:0)α + 3
0.43
We note that, in order to compute the MMSE above, we did not need to compute E[X|Y ] (which is needed by the classical representation of the MMSE), but only the marginal pdf fY , which can be done by simple computations.
Moreover, we also highlight that the MMSE above is in closed form, and this expression highlights that the MMSE only depends on the parameters of the gamma distribution.
Consequently, under the assumption that X is subGaussian, the random variable fσN (X, Z) is bounded by an integrable random variable, and we can apply the dominate convergence theorem to exchange the limit and the expectation and arrive at
REFERENCES [1] E. Weinstein and A. J. Weiss, “A general class of lower bounds in parameter estimation,” IEEE Transactions on Information Theory, vol.
参考 [1] e. weinstein と a. j. weiss, “a general class of lower bounds in parameter estimation”, ieee transactions on information theory, vol. を参照。
0.67
34, no. 2, pp. 338–342, 1988.
34, No. 2, pp. 338–342, 1988。
0.47
[2] H. V. T. K. BELL and Z. Tian, Detection, Estimation, and Modulation
[2] H. V. T. K. BELL and Z. Tian, Detection, Estimation, and Modulation
0.47
Theory Part I. John Wiley & Sons, New York, 1968.
理論第1部。 ジョン・ワイリー&サンズ(John Wiley & Sons)、1968年。
0.63
[3] B. Bobrovsky and M. Zakai, “A lower bound on the estimation error for certain diffusion processes,” IEEE Transactions on Information Theory, vol.
[3] b. bobrovsky and m. zakai, “a lower bound on the estimation error for certain diffusion processes”. ieee transactions on information theory, vol. (英語) 訳抜け防止モード: 3 ] b. bobrovsky, m. zakai, "特定の拡散過程に対する推定誤差の下限" ieee transactions on information theory, vol. を参照。
0.77
22, no. 1, pp. 45–52, 1976.
22 no. 1 pp. 45-52 1976年。
0.74
[4] E. Barankin, “Locally best unbiased estimates,” The Annals of Mathe-
E. Barankin, “Locally best unbiased estimates” The Annals of Mathe-
0.35
matical Statistics, vol.
matical Statistics, vol. (英語)
0.65
20, no. 4, pp. 477–501, 1949.
第20巻、第4巻、1949年、p.477-501。
0.48
[5] B.
B (複数形 Bs)
0.45
-Z. Bobrovsky, E. Mayer-Wolf, and M. Zakai, “Some classes of global
-Z。 ボブロフスキー, e. mayer-wolf, m. zakai, 「世界のいくつかの階級」
0.49
Cram´er-Rao bounds,” The Annals of Statistics, pp. 1421–1438, 1987.
Cram ́er-Rao bounds”. The Annals of Statistics, pp. 1421–1438, 1987.
0.42
[6] J. Ziv and M. Zakai, “Some lower bounds on signal parameter estimation,” IEEE Transactions on Information Theory, vol.
6] j. ziv, m. zakai, “信号パラメータ推定のいくつかの下限”, ieee transactions on information theory, vol。 訳抜け防止モード: [6 ]J. Ziv と M. Zakai, “いくつかの信号パラメータ推定の下位境界” IEEE Transactions on Information Theory, vol。
0.87
15, no. 3, pp. 386–391, 1969.
15, No. 3, pp. 386–391, 1969。
0.95
[7] L. P. Seidman, “Performance limitations and error calculations for parameter estimation,” Proceedings of the IEEE, vol.
7] l. p. seidman, “performance limits and error calculations for parameter estimation” ieee, vol. の議事録。
0.73
58, no. 5, pp. 644– 652, 1970.
58, No. 5, pp. 644– 652, 1970。
0.88
)
)
0.43
英語(論文から抽出)
日本語訳
スコア
[8] D. Chazan, M. Zakai, and J. Ziv, “Improved lower bounds on signal parameter estimation,” IEEE Transactions on Information Theory, vol.
D. Chazan, M. Zakai, J. Ziv, “Improved lower bounds on signal parameters Estimation”, IEEE Transactions on Information Theory, Vol. 訳抜け防止モード: [8]D.Chazan、M.Zkai、J.Ziv 信号パラメータ推定における下限の改善, IEEE Transactions on Information Theory, vol。
0.73
21, no. 1, pp. 90–93, 1975.
21巻1頁、p.90-93、1975年。
0.62
[33] Y. Wu and S. Verd´u, “Optimal phase transitions in compressed sensing,” IEEE Transactions on Information Theory, vol.
IEEE Transactions on Information Theory, vol. [33] Y. Wu and S. Verd ́u, “Optimal phase transitions in compressed sensor”. IEEE Transactions on Information Theory. 訳抜け防止モード: [33 ] Y. Wu, S. Verd ́u, “圧縮センシングにおける最適相転移” IEEE Transactions on Information Theory, vol。
0.87
58, no. 10, pp. 6241– 6263, 2012.
58, no. 10, pp. 6241– 6263, 2012 年。
0.89
[34] D. L. Donoho and J. Tanner, “Counting the faces of randomly-projected hypercubes and orthants, with applications,” Discrete & Computational Geometry, vol.
34] d. l. donoho氏とj. tanner氏は、"ランダムに投影された超キューブとオルタントの顔と応用を数えています。 訳抜け防止モード: D. L. DonohoとJ. Tannerは、”ランダムに投影されたハイパーキューブとオーサントの顔を数える”。 アプリケーション , ” Discrete & Computational Geometry , vol。
0.72
43, no. 3, pp. 522–541, 2010.
43, No. 3, pp. 522–541, 2010
0.44
[35] D. L. Donoho, A. Maleki, and A. Montanari, “Message-passing algorithms for compressed sensing,” Proceedings of the National Academy of Sciences, vol.
35] d. l. donoho, a. maleki, a. montanari, "message-passing algorithms for compression sensing", proceedings of the national academy of sciences, vol. (英語) 訳抜け防止モード: [35 ]D. L. Donoho、A. Maleki、A. Montanari 『メッセージ - 圧縮センシングのための通過アルゴリズム』国立科学アカデミー紀要, Vol.
0.69
106, no. 45, pp. 18 914–18 919, 2009.
106, No. 45, pp. 18 914–18 919, 2009
0.43
[36] D. Guo, D. Baron, and S. Shamai, “A single-letter characterization of optimal noisy compressed sensing,” in 2009 47th Annual Allerton Conference on Communication, Control, and Computing (Allerton).
[36] d. guo氏、d. baron氏、s. shamai氏は2009年の第47回年次allerton conference on communication, control, and computing (allerton)で、"最適なノイズ圧縮センシングの単一レター特性"を発表した。 訳抜け防止モード: [36 ]D・グオ、D・バロン、S・シャマイ。 最適な雑音圧縮センシングの単一文字による特徴付け」 2009年 47th Annual Allerton Conference on Communication, Control, and Computing (Allerton )
0.70
IEEE, 2009, pp. 52–59.
2009年、p.52-59。
0.56
[37] W. Yihong and S. Verd´u, “MMSE dimension,” IEEE Transactions on
[37]W.Yihong and S. Verd ́u, “MMSE dimension”, IEEE Transactions on
0.48
Information Theory, vol. 57, no. 8, pp. 4857–4879, 2011.
情報理論、巻。 57, no. 8, pp. 4857-4879, 2011 頁。
0.71
[9] S. Bellini and G. Tartara, “Bounds on error in signal parameter estimation,” IEEE Transactions on Communications, vol.
9] s. bellini と g. tartara, “bounds on error in signal parameter estimation”, ieee transactions on communications, vol. 。
0.73
22, no. 3, pp. 340–342, 1974.
22 no. 3, pp. 340–342, 1974。
0.88
[10] K. L. Bell, Y. Steinberg, Y. Ephraim, and H. L. Van Trees, “Extended Ziv-Zakai lower bound for vector parameter estimation,” IEEE Transactions on Information Theory, vol.
[10] K. L. Bell, Y. Steinberg, Y. Ephraim, H. L. Van Trees, “Extended Ziv-Zakai lower bound for vector parameters Estimation”, IEEE Transactions on Information Theory, vol。 訳抜け防止モード: [10 ] K. L. Bell, Y. Steinberg, Y. Ephraim そして H. L. Van Trees は “Extended Ziv - Zakai lower bound for vector parameters Estimation” と言う。 IEEE Transactions on Information Theory, vol。
0.90
43, no. 2, pp. 624–637, 1997.
43, no. 2, pp. 624–637, 1997。
0.47
[11] A. Dytso, M. Fauß, A. M. Zoubir, and H. V. Poor, “MMSE bounds for additive noise channels under Kullback–Leibler divergence constraints on the input distribution,” IEEE Transactions on Signal Processing, vol.
A. Dytso, M. Fauß, A. M. Zoubir, H. V. Poor, “MMSE bounds for additive noise channel under the Kullback–Leibler divergence constraints on the input distribution”, IEEE Transactions on Signal Processing, Vol. 訳抜け防止モード: [11] A. Dytso, M. Fauß, A. M. Zoubir. そしてH.V. Poor氏は,“MMSEはKullbackの下で付加的なノイズチャネルのバウンダリ – 入力分布に対するLeibler分散制約”である。 IEEE Transactions on Signal Processing , vol。
0.72
67, no. 24, pp. 6352–6367, 2019.
67, No. 24, pp. 6352-6367, 2019。
0.46
[12] M. Raginsky and I. Sason, “Concentration of measure inequalities in information theory, communications and coding,” arXiv preprint arXiv:1212.4663, 2012.
M. Raginsky, I. Sason, “Concentration of measure inequalities in information theory, communication and coding” arXiv preprint arXiv:1212.4663, 2012 訳抜け防止モード: [12 ]M. Raginsky,I. Sason, “情報理論、通信、符号化における測定の不平等の集中”。 arXiv preprint arXiv:1212.4663 , 2012
0.76
[13] S. L. Fong and V. Y. Tan, “A proof of the strong converse theorem for Gaussian broadcast channels via the Gaussian poincar´e inequality,” IEEE Transactions on Information Theory, vol.
13] s. l. fong and v. y. tan, “a proof of the strongverse theorem for gaussian broadcast channels via the gaussian poincar ́e inequality”. ieee transactions on information theory. (英語) 訳抜け防止モード: [13 ] S. L. Fong と V. Y. Tan は「ガウスのポインカー不等式によるガウス放送の強い逆定理の証明」と述べている。 IEEE Transactions on Information Theory, vol。
0.73
63, no. 12, pp. 7737– 7746, 2017.
63, no. 12 pp. 7737– 7746, 2017 頁。
0.82
[14] Y. Polyanskiy and S. Verd´u, “Empirical distribution of good channel codes with nonvanishing error probability,” IEEE Transactions on Information theory, vol.
14] y. polyanskiy と s. verd ́u, “善良なチャネル符号の非有望なエラー確率によるempirical distribution of good channel codes with nonvanishing error probability”, ieee transactions on information theory, vol.)。 訳抜け防止モード: [14 ] Y. Polyanskiy と S. Verd ́u は,「不要なエラー確率を持つ良チャネル符号の実証分布」である。 IEEE Transactions on Information theory, vol。
0.84
60, no. 1, pp. 5–21, 2013.
60, no. 1, pp. 5-21, 2013年。
0.88
[15] W. Hachem, O. Khorunzhiy, P. Loubaton, J. Najim, and L. Pastur, “A new approach for mutual information analysis of large dimensional multi-antenna channels,” IEEE Transactions on Information Theory, vol.
15] w. hachem氏、o. khorunzhiy氏、p. loubaton氏、j. najim氏、l. pastur氏は、"大規模マルチアンテナチャネルの相互情報分析のための新しいアプローチ"であると述べている。 訳抜け防止モード: [15 ]W. Hachem, O. Khorunzhiy, P. Loubaton, J. Najim, L. Pastur, “大規模マルチアンテナチャネルの相互情報分析のための新しいアプローチ” IEEE Transactions on Information Theory, vol。
0.84
54, no. 9, pp. 3987–4004, 2008.
54, No. 9, pp. 3987–4004, 2008。
0.47
[16] D. Bakry, F. Barthe, P. Cattiaux, and A. Guillin, “A simple proof of the poincar´e inequality for a large class of probability measures,” Electronic Communications in Probability, vol.
16] d. bakry, f. barthe, p. cattiaux, a. guillin, “a simple proof of the poincar 'e inequality for a large class of probability measures”, electronic communications in probability, vol。 訳抜け防止モード: 16] D. Bakry, F. Barthe, P. Cattiaux, A. Guillinは曰く、“多種多様な確率測度に対するポインカーの不等式の簡単な証明”だ。 確率における電子通信 , vol。
0.78
13, pp. 60–66, 2008.
13, pp. 60-66, 2008。
0.85
[17] A. D. Sarwate and A. G. Dimakis, “The impact of mobility on gossip algorithms,” IEEE Transactions on Information Theory, vol.
17] a. d. sarwate, a. g. dimakis, “the impact of mobility on gossip algorithms”, ieee transactions on information theory, vol. ^ モバイルがゴシップアルゴリズムに与える影響。
0.67
58, no. 3, pp. 1731–1742, 2012.
58, No. 3, pp. 1731-1742, 2012
0.39
[18] A. Schlichting, “Poincar´e and log–sobolev inequalities for mixtures,”
18] a. schlichting, "poincar ́e and log-sobolev inequality for mixtures"
0.41
Entropy, vol.
エントロピー、Vol。
0.58
21, no. 1, p. 89, 2019.
21 no. 1 p. 89, 2019。
0.74
[19] A. Vidotto, “An improved second-order poincar´e inequality for functionals of gaussian fields,” Journal of Theoretical Probability, vol.
19] a. vidotto, “an improved second-order poincar ́e inequality for functionals of gaussian fields”. journal of theoretical probability, vol. (英語)
0.42
33, no. 1, pp. 396–427, 2020.
33, No. 1, pp. 396–427, 2020。
0.48
[20] S. Chatterjee, “Fluctuations of eigenvalues and second order poincar´e inequalities,” Probability Theory and Related Fields, vol.
20]s. chatterjee, “fluctuations of eigenvalues and second order poincar ́e inequalities”, probability theory and related fields, vol. . (英語) 訳抜け防止モード: 20]s. chatterjee, 「固有値のゆらぎと二階ポインカーの不等式」 確率論と関連する場 , vol.
0.71
143, no. 1, pp. 1–40, 2009.
143, No. 1, pp. 1–40, 2009
0.44
[21] E. Aras, K.
a b [21] E. Aras, K.
0.39
-Y. Lee, A. Pananjady, and T. A. Courtade, “A family of Bayesian Cram´er-Rao bounds, and consequences for log-concave priors,” in 2019 IEEE International Symposium on Information Theory (ISIT).
-y。 Lee, A. Pananjady, T. A. Courtade, “A family of Bayesian Cram ́er-Rao bounds, and results for log-concave priors” in 2019 IEEE International Symposium on Information Theory (ISIT)。 訳抜け防止モード: -y。 Lee, A. Pananjady, and T. A. Courtade, “A family of Bayesian Cram ́er - Rao bounds, 2019 IEEE International Symposium on Information Theory (ISIT ) に収録されている。
0.59
IEEE, 2019, pp. 2699–2703.
IEEE, 2019, pp. 2699-2703。
0.45
[22] H. Robbins, “An empirical Bayes approach to statistics,” in Proc.
procの[22] h. robbinsは“統計に対する経験的ベイズアプローチ”だ。
0.68
Third Berkeley Symp.
3代目 バークレー・シンプ
0.34
Math Statist. Probab.
数学統計学者。 プロバブ
0.43
Citeseer, 1956.
1956年、デビュー。
0.70
[23] R. Esposito, “On a relation between detection and estimation in decision
[23]R. Esposito「意思決定における検出と推定の関係について」
0.80
theory,” Inf. Control, vol.
とinfは言った。 コントロール、vol。
0.52
12, no. 2, pp. 116–120, February 1968.
12 no. 2, pp. 116–120, 1968年2月。
0.88
[24] A. Dytso, H. V. Poor, and S. Shamai, “A general derivative identity for the conditional mean estimator in Gaussian noise and some applications,” arXiv preprint arXiv:2104.01883, 2021.
[24] A. Dytso, H. V. Poor, S. Shamai, “A general derivative identity for the conditional mean estimator in Gaussian noise and some applications” arXiv preprint arXiv:2104.01883, 2021. 訳抜け防止モード: [24 ]A. Dytso、H.V. Poor、S. Shamai 「ガウス雑音における条件平均推定器の一般微分等式とその応用」 arXiv preprint arXiv:2104.01883 , 2021。
0.79
[25] A. Dytso and M. Cardone, “A general derivative identity for the conditional expectation with focus on the exponential family,” arXiv preprint arXiv:2105.05106, 2021.
[25] A. Dytso と M. Cardone, “A general derivative identity for the conditional expectation with focus on the index family” arXiv preprint arXiv:2105.05106, 2021. 訳抜け防止モード: 25 ] a. dytso と m. cardone は "指数関数族に焦点をあてた条件付き期待に対する一般的な微分id" である。 arxiv プレプリント arxiv:2105.05106 、2021 。
0.58
[26] S. Boucheron, G. Lugosi, and P. Massart, Concentration Inequalities: A Nonasymptotic Theory of Independence.
S. Boucheron, G. Lugosi, and P. Massart, concentration inequalities: A Nonasymsymotic Theory of Independence 訳抜け防止モード: 26] s. boucheron, g. lugosi, p. massart 濃度不等式 : 独立の非漸近理論
0.62
Oxford university press, 2013.
オックスフォード大学出版局、2013年。
0.67
[27] D. Bakry and M. ´Emery, “Diffusions hypercontractives,” in Seminaire
[27]D. Bakry and M. Émery, “Diffusions hypercontractives” in Seminaire
0.43
de probabilit´es XIX 1983/84.
1983/84シーズンにデビュー。
0.38
Springer, 1985, pp. 177–206.
春田、1985年、p.177-206。
0.58
[28] S. Bobkov and M. Ledoux, “Poincar´e’s inequalities and Talagrand’s concentration phenomenon for the exponential distribution,” Probability Theory and Related Fields, vol.
28] s. bobkov と m. ledoux, “poincar ́e の不等式と talagrand の指数分布の集中現象”, 確率論と関連する分野, vol。 訳抜け防止モード: [28 ]S. Bobkov と M. Ledoux, “Poincar ́e の不等式” そして、指数分布に対するタラグランドの濃度現象「確率論」 および関連フィールド。
0.85
107, no. 3, pp. 383–400, 1997.
107, No. 3, pp. 383–400, 1997。
0.93
[29] Y. Polyanskiy, H. V. Poor, and S. Verd´u, “Channel coding rate in the finite blocklength regime,” IEEE Transactions on Information Theory, vol.
29] y. polyanskiy, h. v. poor, and s. verd ́u, “channel coding rate in the finite blocklength regime”, ieee transactions on information theory, vol。 訳抜け防止モード: [29 ] Y. Polyanskiy, H. V. Poor, S. Verd ́u. 有限ブロック長系におけるチャネル符号化率, IEEE Transactions on Information Theory, vol。
0.89
56, no. 5, pp. 2307–2359, 2010.
56, No. 5, pp. 2307–2359, 2010
0.44
[30] D. Guo, Y. Wu, S. Shamai (Shitz), and S. Verd´u, “Estimation in Gaussian noise: properties of the minimum mean-square error,” IEEE Transactions on Information Theory, vol.
D. Guo, Y. Wu, S. Shamai (Shitz), and S. Verd ́u, “Estimation in Gaussian noise: properties of the least mean-square error”, IEEE Transactions on Information Theory, Vol. 訳抜け防止モード: [30 ]D.Guo,Y.Wu,S.Shamai( シッツ) S. Verd ́u は「ガウス雑音の推定 : 最小平均-二乗誤差の性質」である。 IEEE Transactions on Information Theory, vol。
0.82
57, no. 4, pp. 2371–2385, 2011.
57, No. 4, pp. 2371–2385, 2011
0.39
[31] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
統計信号処理の基礎-[31]S.M.ケイ, 推定
0.79
Theory. Prentice-Hall, Inc., 1993.
理論 Prentice-Hall, Inc., 1993年。
0.63
[32] D. Guo, S. Shamai (Shitz), and S. Verd´u, “Mutual information and minimum mean-square error in Gaussian channels,” IEEE Transactions on Information Theory, vol.
IEEE Transactions on Information Theory, vol.[32]D. Guo, S. Shamai (Shitz), S. Verd ́u, “Kutual information and least mean-square error in Gaussian channel”. IEEE Transactions on Information Theory. 訳抜け防止モード: [32 ]D.Guo, S. Shamai (Shitz ) とS. Verd ́uは言う。「ガウス流路における相互情報と最小平均 - 正方誤差」。 IEEE Transactions on Information Theory, vol。