📊 statistics

On the Expectation of the Local-to-Zero Cross-Validated Log Likelihood Criterion for Bandwidth Selection in Kernel Spectral Estimation

This paper establishes that the expectation of a key term in the Taylor series expansion for a local-to-zero cross-validated log-likelihood criterion ( $\mbox{CVLL}_c$ ) converges to the asymptotic mean squared error of the spectral estimator at zero frequency when $4/5 < c < 1$ , thereby providing theoretical justification for using this local criterion in HAC standard error estimation.

Original authors: Meng-Chen Hsieh, Clifford Hurvich

Published 2026-07-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Meng-Chen Hsieh, Clifford Hurvich

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but the clues you have are a bit fuzzy. In the world of statistics, this "fuzziness" often comes from time series data—think of stock prices, weather patterns, or heartbeats recorded over time. These data points aren't independent; what happens now often depends on what happened a moment ago. To understand the true story behind the noise, statisticians use a tool called a "spectral estimator." You can think of this as a special pair of glasses that helps you see the hidden rhythms or frequencies in the data.

However, these glasses need to be focused just right. The "focus knob" is called the bandwidth. If you turn it too far one way, the picture is too blurry to see any details. If you turn it too far the other way, the picture is so grainy with static that you can't see the shape at all. Finding the perfect setting is crucial, especially when you are looking at the very beginning of the frequency spectrum (zero frequency), which tells us about the long-term stability of the data. This is vital for calculating "standard errors," which are the confidence intervals researchers use to say, "We are pretty sure this result isn't just a fluke." For decades, statisticians have used a clever trick called Cross-Validated Log-Likelihood (CVLL) to find this perfect focus. It works by pretending to leave out one piece of data at a time, seeing how well the rest of the data predicts it, and adjusting the knob until the prediction is best. But here's the catch: the old trick was designed to look at the entire range of frequencies, like scanning a whole radio dial. The new problem is that sometimes we only care about the very bottom of the dial, near zero.

This paper, written by Meng-Chen Hsieh and Clifford Hurvich, tackles a specific puzzle: Does the old "leave-one-out" trick still work when we shrink our view to focus only on the frequencies near zero? The authors introduce a modified version of the trick, called CVLLc, which only looks at a small slice of frequencies (from 1 up to a fraction $c$ of the total, where $c$ is between 0 and 1). They wanted to know if minimizing this new, local score would actually lead them to the best possible bandwidth for estimating the long-term variance.

The authors dive deep into the math to see what happens on average when you use this new local method. They break down a complex formula into smaller pieces, looking at the "expected" behavior of the most important part. Their main finding is a bit like discovering that a specific type of magnifying glass only works if you hold it at a very specific distance. They prove that if the fraction of frequencies you look at ( $c$ ) is large enough—specifically, if $c$ is greater than $4/5$ (or 0.8)—then the average performance of this new local method converges to the true "mean squared error." In plain English, this means that if you look at the bottom 80% or more of the available frequencies, the local method is mathematically justified and will lead you to the correct bandwidth setting.

However, the paper also draws a clear line in the sand. The authors explain that their results cannot be derived from the existing literature on the global method. The old math relied on a property called Parseval's formula, which acts like a universal energy conservation law for the whole radio dial. But when you zoom in on just the zero-frequency end of the dial, that law doesn't hold up in the same way. The authors show that you can't just assume the old rules apply to the new, local version; you have to do the heavy lifting of proving it from scratch. They also note that while they focused on the "expectation" (the average outcome) of the key term, they are confident that the other messy terms in the equation become negligible as the sample size grows, though they leave the full proof of those specific terms for future work.

So, what does this mean for the real world? It provides a solid theoretical foundation for using this local-to-zero method when researchers need to estimate standard errors for data with complex patterns (heteroskedasticity and autocorrelation). It tells us that as long as we don't cut off too much of the frequency spectrum (keeping $c > 4/5$ ), we can trust this local cross-validation method to tune our statistical glasses correctly. It's a reassuring confirmation that a clever shortcut, when applied with the right boundaries, is actually a reliable path to the truth.

Technical Summary: Expectation of the Local-to-Zero Cross-Validated Log-Likelihood Criterion

Problem Statement
The paper addresses the problem of data-driven bandwidth selection for kernel spectral estimators specifically at zero frequency ( $\omega = 0$ ). This estimation is critical for Heteroskedasticity and Autocorrelation Consistent (HAC) standard error estimation, where the long-run variance of a time series is proportional to the spectral density at zero, $f(0)$ .

While global bandwidth selection criteria exist, such as the Cross-Validated Log-Likelihood (CVLL) proposed by Beltrao and Bloomfield (1987) and analyzed by Robinson (1991), these methods sum over all Fourier frequencies in $(0, \pi)$ . The authors note that existing theoretical justifications for CVLL rely on the global nature of the sum, specifically exploiting Parseval's formula. However, HAC estimation focuses on a "local-to-zero" neighborhood. Xu and Hurvich (2026) previously proposed a local-to-zero version of the criterion, denoted $CVLL_c$ , which sums over Fourier frequencies from $1$ to $\tilde{n} = nc$ (where $0 < c < 1$ ). The theoretical gap this paper fills is the lack of rigorous justification for $CVLL_c$ ; specifically, it is unknown whether the expectation of this local criterion converges to the asymptotic mean squared error (MSE) of the spectral estimator at zero frequency, as the global results do not directly apply.

Methodology
The authors analyze the behavior of the local criterion $CVLL_c(m)$ , defined as:
$CVLL_c(m) = \frac{1}{\tilde{n}} \sum_{j=1}^{\tilde{n}} \left[ \hat{f}_{m}^{(j)}(\omega_j) + \frac{I(\omega_j)}{\hat{f}_{m}^{(j)}(\omega_j)} \right]$
where $\hat{f}_{m}^{(j)}$ is the leave-one-out spectral density estimate at frequency $\omega_j$ , and $I(\omega_j)$ is the periodogram.

The analysis proceeds via a Taylor series expansion of the difference between the local log-likelihood $\tilde{L}(\tau)$ and a baseline $\tilde{L}$ (where $\tau$ is the bandwidth tuning parameter such that $m = \tau n^{1/5}$ ). The expansion yields several terms:

A linear term involving $(\hat{f}/f - 1)(I/f - 1)$ .
A quadratic term involving $(I/f - 1)(\hat{f}/f - 1)^2$ .
A quadratic term involving $(\hat{f}/f - 1)^2$ .
Higher-order remainder terms.

The authors focus on the expectation of the third term (Eq. 0.18 in the text), scaled by $n^{4/5-c}$ . They argue that for Gaussian white noise, the expectations of the first two terms are exactly zero due to independence. While they do not prove the negligibility of these terms for general processes, they posit that under suitable regularity conditions, they are negligible, and the primary behavior is driven by the third term.

The core of the proof involves decomposing the estimator error $(\hat{f}_j^{(j)}/f_j - 1)$ into components related to the lag-window estimate and the periodogram of the innovations. The authors utilize properties of Parzen's lag window function and the asymptotic behavior of the sample autocovariances.

Key Results
The paper establishes the following main theorem:
Let $m = \tau n^{1/5}$ and $\tilde{n} = n^c$ with $4/5 < c < 1$ . As $n \to \infty$ , the expectation of the key quadratic term in the Taylor expansion converges to the asymptotic mean squared percentage error of the spectral estimator at zero frequency:
$\frac{1}{2} \left[ n^{4/5-c} \sum_{j=1}^{\tilde{n}} E \left( \frac{\hat{f}_j^{(j)}}{f_j} - 1 \right)^2 \right] \to c_0(\tau)$
where $c_0(\tau)$ is the asymptotic MSE function composed of a variance term ( $\tau \kappa$ ) and a squared bias term ( $\tau^{-4} [h^{(2)} f^{(2)}(0)/f(0)]^2$ ).

The proof relies on several lemmas:

Lemma 0.2: Demonstrates that the squared bias component of the expectation converges uniformly to the theoretical bias term $b_0(\tau)$ .
Lemma 0.3 & Corollary 0.4: Establish bounds on the Dirichlet kernel sums over the local frequency range, showing that certain cross-terms vanish as $n \to \infty$ provided $c > 4/5$ .
Lemma 0.5: Proves that the variance component of the expectation converges to $\tau \kappa$ , where $\kappa = \int k^2(x) dx$ .

Crucially, the authors show that the condition $c > 4/5$ is necessary for the convergence of the variance term, distinguishing the local case from the global case where such a restriction on the summation limit does not apply in the same manner.

Significance and Claims
The paper claims to provide "some justification" for the use of the local-to-zero CVLL criterion ( $CVLL_c$ ) in HAC standard error estimation. By demonstrating that the expectation of the dominant term in the local criterion converges to the asymptotic MSE of the estimator at zero frequency, the authors validate the theoretical underpinning of minimizing $CVLL_c$ to select the optimal bandwidth.

The authors explicitly state that their results do not follow from existing literature on global CVLL (e.g., Robinson, 1991) because the global proofs rely on Parseval's formula, which does not hold for sums restricted to a neighborhood of zero frequency. This work therefore fills a specific theoretical gap regarding the asymptotic properties of local spectral estimation criteria. The paper remains modest in its claims, noting that it focuses on the expectation of a specific term and does not fully prove the negligibility of the other Taylor expansion terms for non-Gaussian processes, though it suggests this is likely true under regularity conditions.

Technical Summary: Expectation of the Local-to-Zero Cross-Validated Log-Likelihood Criterion

More like this