正则化逆的一致收敛性：关于||(ARₑ-I)m||的技术问询

Alright, let’s break this down—this is a core question in regularization theory for ill-posed problems, so let’s unpack what we can conclude about the norm $|(AR_\varepsilon - I)m|$ given your setup.

First, let’s restate the key assumptions to keep things clear:

• $A$ is a bounded, compact, injective linear operator (we’ll assume Hilbert spaces for concreteness, since most regularization frameworks live here, but we’ll touch on Banach spaces too).
• $m \in \text{Ran}(A)$, so there’s a unique $f$ such that $Af = m$ (injectivity guarantees uniqueness).
• $R_\varepsilon$ is a regularized inverse of $A$ with parameter $\varepsilon > 0$, e.g., Tikhonov or spectral truncation.

Step 1: Rewrite the Norm for Simplification #

First, substitute $m = Af$ into the expression we’re studying:

||(ARₑ - I)m|| = ||ARₑAf - Af|| = ||A(RₑAf - f)||

Since $A$ is bounded, this immediately tells us the norm is bounded by $|A| \cdot |RₑAf - f|$. But we can get much sharper results by looking at specific regularization methods.

2. Key Properties for Common Regularization Inverses #

Let’s dive into the two most widely used regularized inverses:

Spectral Truncation Inverse #

For compact operators in Hilbert spaces, we can use the singular value decomposition (SVD) of $A$: $A = \sum_{n=1}^\infty \lambda_n u_n v_n^*$, where $\lambda_n > 0$ are the singular values (decreasing to 0, since $A$ is compact), and ${u_n}, {v_n}$ are orthonormal bases.

The spectral truncation regularized inverse $R_\varepsilon$ is defined as:

Rₑy = sum_{λₙ > ε} (1/λₙ) v_n ⟨y, u_n⟩

Now compute $AR_\varepsilon m$:

• Since $m = Af = \sum_{n=1}^\infty λ_n ⟨f, v_n⟩ u_n$, substituting into $R_\varepsilon m$ gives us $\sum_{λₙ > ε} ⟨f, v_n⟩ v_n$.
• Applying $A$ to this sum gives $AR_\varepsilon m = \sum_{λₙ > ε} λ_n ⟨f, v_n⟩ u_n$.

Subtract $m$ to get:

(ARₑ - I)m = sum_{λₙ ≤ ε} λ_n ⟨f, v_n⟩ u_n

Taking the norm squared:

||(ARₑ - I)m||² = sum_{λₙ ≤ ε} λₙ² |⟨f, v_n⟩|²

Key Properties:

• Convergence as $\varepsilon \to 0$: The norm tends to 0, because the sum only includes terms with $\lambda_n ≤ \varepsilon$, which vanishes as $\varepsilon$ shrinks (and the original series $\sum λₙ² |⟨f, v_n⟩|² = |m|²$ is convergent).
• Convergence Rate: Depends on the "smoothness" of $f$. If $f \in \text{Ran}(A^k)$ (i.e., $f = A^k g$ for some $g$), then $|⟨f, v_n⟩| = λ_n^k |⟨g, v_n⟩|$. Substituting this in gives:

  ||(ARₑ - I)m||² ≤ ε² sum_{λₙ ≤ ε} λₙ^{2(k-1)} |⟨g, v_n⟩|² ≤ ε^{2k} ||g||²
  

  So the convergence rate is $O(\varepsilon^k)$—smoother $f$ (higher $k$) leads to faster convergence.

Tikhonov Regularization Inverse #

The Tikhonov regularized inverse is defined as:

Rₑy = (A^*A + εI)^{-1} A^* y

Again, using the SVD of $A$, let $B = A^A = \sum_{n=1}^\infty λ_n² v_n v_n^$. Then:

(A^*A + εI)^{-1} A^*A = sum_{n=1}^\infty (λ_n² / (λ_n² + ε)) v_n v_n^*

Substitute $m = Af$ into $AR_\varepsilon m$:

ARₑm = A sum_{n=1}^\infty (λ_n² / (λ_n² + ε)) ⟨f, v_n⟩ v_n = sum_{n=1}^\infty (λ_n³ / (λ_n² + ε)) ⟨f, v_n⟩ u_n

Subtract $m$ to get:

(ARₑ - I)m = sum_{n=1}^\infty \left( \frac{λ_n³}{λ_n² + ε} - λ_n \right) ⟨f, v_n⟩ u_n = sum_{n=1}^\infty \left( \frac{-ε λ_n}{λ_n² + ε} \right) ⟨f, v_n⟩ u_n

Taking the norm squared:

||(ARₑ - I)m||² = sum_{n=1}^\infty \left( \frac{ε² λ_n²}{(λ_n² + ε)^2} \right) |⟨f, v_n⟩|²

Key Properties:

• Convergence as $\varepsilon \to 0$: The norm tends to 0, since each term in the series goes to 0 (as $\varepsilon/(λ_n² + ε) \to 0$) and the series is dominated by $\sum λ_n² |⟨f, v_n⟩|² = |m|²$.
• Convergence Rate: Again, depends on $f$’s smoothness. If $f \in \text{Ran}(A^k)$, then $|⟨f, v_n⟩| = λ_n^k |⟨g, v_n⟩|$. Using the inequality $\frac{ε λ_n}{λ_n² + ε} ≤ C ε^{k/(k+1)} λ_n^{k/(k+1)}$ (from optimizing the rational function), we get a convergence rate of $O(\varepsilon^{k/(k+1)})$. This is slower than spectral truncation for the same $k$, but Tikhonov has better stability properties for noisy data.

3. General Banach Space Case #

If we’re working in a Banach space (not just Hilbert), the core convergence result still holds: since $R_\varepsilon$ is a regularized inverse, it satisfies $\lim_{\varepsilon \to 0} R_\varepsilon y = A^{-1} y$ for all $y \in \text{Ran}(A)$. For $m = Af$, this means $R_\varepsilon m = R_\varepsilon Af \to f$ as $\varepsilon \to 0$. Since $A$ is bounded, $|A(R_\varepsilon Af - f)| ≤ |A| \cdot |R_\varepsilon Af - f| \to 0$, so $|(AR_\varepsilon - I)m| \to 0$.

Summary #

To wrap up:

• For any valid regularized inverse $R_\varepsilon$, $|(AR_\varepsilon - I)m| \to 0$ as $\varepsilon \to 0$ (since we’re converging to the true solution’s image).
• The speed of convergence depends on how "smooth" $f$ is (i.e., how quickly its coefficients decay in the singular value expansion of $A$).
• Different regularization methods give different convergence rates: spectral truncation is faster for smooth $f$, while Tikhonov offers better robustness to noise.

内容的提问来源于stack exchange，提问作者SecretlyAnEconomist