Stationary Process Invertibility and the Unilateral Shift Operator

This paper argues that the unilateral shift operator offers a superior framework for analyzing stationary process invertibility by rigorously establishing that for functions in the Wiener algebra, the associated operator is a Toeplitz operator with a norm equal to the function's supremum norm, thereby unifying statistical and algebraic notions of invertibility.

The Gist

The Big Picture: Two Ways to Look at Time

Imagine you are trying to predict the future of a weather pattern, a stock price, or a song. In statistics, we call these Stationary Processes. They are things that behave consistently over time.

For decades, statisticians have used a specific mathematical tool called the Bilateral Shift Operator (let's call it B). Think of B as a magical time machine that can look both backward and forward into infinity. It assumes the process has been running forever and will run forever.

The authors of this paper argue that for a specific problem—Invertibility—this "infinite time machine" is the wrong tool. Instead, we should use the Unilateral Shift Operator (let's call it T). Think of T as a camera that only looks backward. It sees the past and the present, but it has no lens for the future.

The Problem: The "Infinite" Trap

What is Invertibility?
In plain English, invertibility asks: "Can we figure out the original cause (the noise) just by looking at the current result?"

Imagine you have a smoothie (the result) made from a blender.

• Invertible: If you can perfectly reverse the blender's action to get the exact original strawberries and milk back, the process is invertible.
• Non-Invertible: If the blender mashed the fruit so thoroughly that you can't tell which strawberry was which, or if the machine has a "memory" that distorts the past, you can't reverse it.

The Conflict:
When statisticians use the Bilateral Operator (B), they assume the machine has been running forever. Because of this infinite assumption, the math sometimes says a process is "invertible" (reversible) even when it clearly isn't in the real world.

The Analogy:
Imagine a song played on a record player.

• Using B (Infinite): The record player is a magical loop that plays the song forever in both directions. If you try to reverse the song, the math says, "Oh, sure, you can go back to the beginning because the song has always existed."
• Using T (Real World): The record player starts at the beginning of the song. If you try to reverse it past the start, the needle hits the label and stops. The math correctly says, "No, you can't go back further than the start."

The paper argues that for real-world modeling (like predicting tomorrow's weather based on today's), we must use T (the camera that only looks back), not B.

The "Magic" of the Wiener Algebra

The authors dive into a specific mathematical club called the Wiener Algebra (denoted as W).

• Think of this as a "VIP Lounge" for functions. To get in, your function must be "well-behaved" (specifically, the sum of its parts must be finite, known as the $\ell_1$ condition).

The paper proves three major things about this VIP Lounge when using the Unilateral Shift (T):

1. Existence: If you take a "VIP" function and apply it to our time-camera T, the result is a valid, well-defined mathematical object. It doesn't break the math.
2. Isometry (The Perfect Mirror): They prove that the "size" (norm) of the function in the algebra is exactly the same as the "size" of the operator it creates.
   • Analogy: Imagine a shadow puppet show. Usually, the shadow might look bigger or smaller than the hand making it. The authors prove that for these specific functions, the shadow is the exact same size as the hand. No distortion.
3. Identity: They show that this new operator f(T) is actually the same thing as a famous tool called the Toeplitz Operator. It's like discovering that two different names you've been calling your best friend are actually the same person.

Why Does This Matter?

For a long time, textbooks taught that if a process met a strict "sum of parts" rule ($\ell_1$), it was invertible. But nobody knew if that rule was the only way to be invertible, or if it was just a safe, easy rule.

The authors suggest that by switching from the "Infinite Time Machine" (B) to the "Real-World Camera" (T), we might be able to relax those strict rules.

• Current Rule: "You must be perfectly smooth and summable to be reversible."
• New Hope: "Maybe you just need to be 'bounded' (not exploding to infinity) to be reversible."

They didn't solve the whole puzzle in this paper, but they built the foundation (the "rigorous operator theoretic foundation") to try solving it later. They are essentially saying: "Stop using the infinite model for finite problems. Use the unilateral model, and we might find a more flexible way to understand how the past creates the present."

Summary in a Nutshell

• The Old Way: Used a tool that assumes time goes on forever in both directions. It was good for averages, but bad for understanding how to reverse a process.
• The New Way: Use a tool that respects the fact that time starts somewhere and moves forward.
• The Result: By using this new tool, the authors proved that the math works perfectly for a specific class of functions, and they opened the door to making the rules for "reversibility" less strict and more realistic.

Technical Summary

1. Problem Statement

The paper addresses a fundamental disconnect in the literature regarding the modeling of stationary stochastic processes, specifically concerning invertibility.

• Current Practice: Standard time series literature (e.g., Box & Jenkins, Brockwell & Davis, Hamilton) predominantly utilizes the bilateral shift operator ($B$) to model stationary processes. This operator acts on doubly infinite sequences ($n \in \mathbb{Z}$).
• The Conflict: While the bilateral shift is sufficient for calculating statistical parameters (mean, covariance) as argued by Doob, the authors contend it is inappropriate for analyzing invertibility (the ability to represent a process as an autoregressive process based on its past).
• The Specific Issue:
  • Invertibility in time series requires that a process $X_t$ can be expressed as a convergent sum of past values ($X_t = \sum b_n X_{t-n}$).
  • Using the bilateral shift $B$, the transfer function $f(B)$ is often invertible even when the roots of $f(z)$ lie inside the unit circle (because $B^{-1}$ exists). This contradicts the standard time series definition where such processes are non-invertible.
  • The prevailing condition for invertibility is the $\ell_1$ condition (absolute summability of coefficients), which is known to be sufficient but its necessity is unclear.
• Goal: The authors aim to unify the statistical notion of process invertibility with the algebraic invertibility of transfer functions by shifting the mathematical framework from the bilateral operator $B$ to the unilateral shift operator $T$ (acting on singly infinite sequences, $n \in \mathbb{N}$).

2. Methodology

The authors employ Operator Theory and Functional Analysis to reframe the problem.

• Shift Operators:
  • Bilateral Shift ($B$): Defined on a Hilbert space with basis $\{e_n : n \in \mathbb{Z}\}$, where $B(e_n) = e_{n+1}$. It is a unitary operator.
  • Unilateral Shift ($T$): Defined on a Hilbert space with basis $\{e_n : n \in \mathbb{N}\}$, where $T(e_n) = e_{n+1}$. It is a contraction (not unitary, as $T^{-1}$ does not exist).
• Nagy's Dilation Theorem: The authors utilize this theorem, which states that any contraction $T$ on a Hilbert space can be viewed as the compression of a unitary operator $B$ (its dilation) on a larger space. This bridges the gap between the singly infinite (causal) and doubly infinite processes.
• Wiener Algebra ($W$): The analysis is restricted to functions $f(z) = \sum a_n z^n$ where the coefficients $(a_n)$ are in $\ell_1$. This ensures absolute convergence of the power series defining the operators.
• Hardy Spaces ($H^2$) and Toeplitz Operators: The authors map the unilateral shift operator to the multiplication operator on the Hardy space $H^2$ and relate the transfer function operator $f(T)$ to the Toeplitz operator $T_f$.

3. Key Contributions and Results

A. Unification of Invertibility Concepts

The paper argues that the unilateral shift operator $T$ is the mathematically correct tool for invertibility because it respects the causal structure (dependence only on the past).

• Example: For a transfer function $f(z) = 1 - 2z$ (root at $z=0.5$, inside the unit circle):
  • Using $B$: $f(B) = 1 - 2B$ is invertible because $B^{-1}$ exists. This yields a non-causal expansion involving future terms.
  • Using $T$: $f(T) = 1 - 2T$ is not invertible because $T^{-1}$ does not exist. This correctly aligns with the time series definition that the process is non-invertible.

B. Rigorous Operator-Theoretic Foundations

The authors establish three main propositions regarding the transfer function operator $f(T)$ for $f$ in the Wiener algebra ($W$):

1. Existence (Proposition 1):
   For $f \in W$, the power series $f(T) = \sum a_n T^n$ is a well-defined bounded linear operator. The series converges in the operator norm. This is proven by utilizing the spectral theorem on the unitary dilation $B$ and projecting it back to the subspace of the unilateral shift.

2. Isometry (Proposition 2):
   The authors prove that the operator norm of the transfer function equals the supremum norm of the function:
   $$\|f(T)\| = \|f\|_\infty$$
   This result improves upon previous norm inequalities found in the literature. The proof relies on the fact that $f(T)$ is unitarily equivalent to the multiplication operator $M_f$ on the Hardy space $H^2$, and utilizes the properties of the Szegő kernel.

3. Equivalence to Toeplitz Operators (Proposition 3):
   For $f \in W^+$ (causal functions, $a_n=0$ for $n<0$), the operator $f(T)$ is identical to the Toeplitz operator $T_f$:
   $$f(T) = T_f$$
   This provides a rigorous link between the abstract power series definition and the standard Toeplitz operator definition used in harmonic analysis.

C. Critique of the $\ell_1$ Condition

The paper highlights that the standard $\ell_1$ condition (absolute summability) is a sufficient condition for invertibility but questions its necessity. The authors suggest that the condition might be relaxed to require only mean square convergence (associated with $H^\infty$ functions) rather than absolute convergence, though they note this requires more advanced machinery (von Neumann algebras) which is reserved for future work.

4. Significance and Implications

• Theoretical Correction: The paper corrects a subtle but significant theoretical oversight in time series literature. By using the bilateral shift for invertibility, standard texts inadvertently allow for non-causal inverses, which contradicts the physical reality of time series modeling (where the future cannot predict the past).
• Unification: It successfully unifies the statistical concept of "process invertibility" with the algebraic concept of "transfer function invertibility" by adopting the unilateral shift operator.
• Foundation for Future Work: The results provide a rigorous operator-theoretic foundation for extending the definition of invertibility beyond the Wiener algebra ($\ell_1$) to the broader class of $H^\infty$ functions. This could lead to relaxed conditions for invertibility in practical time series analysis, allowing for a wider class of models to be considered invertible under mean-square convergence criteria.
• Methodological Shift: The paper advocates for a shift in pedagogical and research approaches, moving away from the "convenience" of the bilateral shift for all problems and reserving it for statistical parameter estimation, while using the unilateral shift for structural and algebraic questions like causality and invertibility.

Conclusion

The authors demonstrate that the unilateral shift operator $T$ is the natural and necessary framework for analyzing the invertibility of stationary processes. By proving that $f(T)$ is well-defined, isometric, and equivalent to the Toeplitz operator for functions in the Wiener algebra, they resolve the algebraic inconsistencies present in standard time series texts and pave the way for generalizing invertibility conditions to $H^\infty$ transfer functions.