1 Motivation and Basic Setup
1.1 Covariance in dependent sequences
Many limit theorems are straightforward for independent or nearly independent observations, but dependence can alter both the growth rate of variances and the form of limiting distributions. A central role is played by covariances across indices: they quantify how fluctuations at one location in the sequence co-move with fluctuations elsewhere. Covariance summability asks whether these co-movements, accumulated across increasing separations, remain tame enough for asymptotic statements to mirror the independent case.
1.2 Common indexing conventions (lags, spatial indices)
Covariance summability is usually formulated for a single indexing parameter (often time) using lags, or for multi-indexed settings (e.g., spatial sites) using separation vectors. For time-indexed sequences, a standard stationary setup uses covariances of the form \[ \gamma(k)=\mathrm{Cov}(X_0,X_k). \] For spatial indices, one often sums over neighborhoods or over shells of increasing radius, effectively asking whether covariances decay sufficiently as distance grows.
1.3 The role of summability in variance control
Consider partial sums \(S_n=\sum_{t=1}^n X_t\). The variance of \(S_n\) contains not only the individual variances but also all cross-covariances between different times. Summability conditions control whether the total contribution of these cross-terms grows linearly with \(n\) (the “well-behaved” regime) or grows faster (indicating strong long-range dependence). When the relevant covariance series converges, one can often define a finite asymptotic variance and use it in both law of large numbers and central limit theorem statements.
2 Definitions of Covariance Summability
2.1 Absolute summability of lag covariances
A canonical notion is absolute summability of lag covariances. For a stationary sequence, \[
| \sum_{k=-\infty}^{\infty} | \gamma(k) | < \infty. |
|---|
\]
| Because \(\gamma(-k)=\gamma(k)\) under standard symmetry for real-valued processes, this is equivalent to \(\sum_{k=0}^{\infty} | \gamma(k) | <\infty\). This condition ensures that long-lag dependence contributes a finite amount to aggregated variance, preventing divergence from accumulation of small but numerous correlations. |
|---|
2.1.1 Covariance series across integer lags
The “across integer lags” viewpoint emphasizes the discrete structure: one sums \(\gamma(k)\) over \(k\in\mathbb{Z}\). Absolute summability strengthens mere convergence of \(\sum_k \gamma(k)\) by controlling cancellation. Without absolute control, alternating signs can mask the magnitude of dependence, and variance behavior may be harder to guarantee.
2.2 Square-summability and related norms
Another family of conditions uses norms rather than direct absolute summation. A typical requirement is square-summability, \[ \sum_{k=-\infty}^{\infty} \gamma(k)^2 < \infty, \] or related block/neighborhood criteria that bound integrated covariance strength over finite regions. These conditions often arise in proving bounds via Hilbert-space methods or in establishing existence of certain spectral objects.
2.2.1 Summability over blocks or neighborhoods
For non-time-homogeneous or multi-dimensional index sets, summability may be formulated over regions: \[
| \sum_{\text{distance}(i,j)\le r} | \mathrm{Cov}(X_i,X_j) | , |
|---|
\] or in terms of how such totals grow with \(r\). The key is that covariance mass in increasingly large neighborhoods should not grow faster than linearly in the aggregation scale.
2.3 Summability of covariances under different scalings
In some settings, covariances are rescaled before summation, reflecting growth patterns of the index set. For example, one may consider \[
| \sum_{k=1}^\infty w(k)\, | \gamma(k) | < \infty |
|---|
\] for a weight \(w(k)\) that decays or grows with \(k\). Weighted versions are useful when the variance of \(S_n\) is known to scale in a nonstandard way, or when one studies averages over more complex geometries.
3 Variance of Partial Sums and Convergence
3.1 Expressing Var(S_n) via covariances
For \(S_n=\sum_{t=1}^n X_t\), \[ \mathrm{Var}(S_n)=\sum_{t=1}^n \mathrm{Var}(X_t)\;+\;2\sum_{1\le s<t\le n}\mathrm{Cov}(X_s,X_t). \] Under stationarity with \(\gamma(k)=\mathrm{Cov}(X_0,X_k)\), this becomes \[ \mathrm{Var}(S_n)=n\gamma(0)+2\sum_{k=1}^{n-1}(n-k)\gamma(k). \] Thus, the asymptotic behavior hinges on how the terms \((n-k)\gamma(k)\) accumulate as \(n\) grows.
3.2 When normalized variance converges
| Define the normalized variance \(n^{-1}\mathrm{Var}(S_n)\). If \(\sum_{k=1}^\infty | \gamma(k) | <\infty\), then |
|---|
\[ \frac{1}{n}\mathrm{Var}(S_n)\to \gamma(0)+2\sum_{k=1}^{\infty}\gamma(k), \] provided \(\sum_{k\ge 1}\gamma(k)\) is finite as an ordinary series (absolute summability is a sufficient condition). When covariances decay too slowly, the sum may fail to converge and normalized variance may diverge or oscillate in scale.
3.3 Interpreting the limiting (asymptotic) variance
3.3.1 Absolute-summability implications
Under absolute summability, the “effective dependence” accumulates to a finite total, yielding a well-defined asymptotic variance \[ \sigma^2=\gamma(0)+2\sum_{k=1}^{\infty}\gamma(k). \] This constant often becomes the variance parameter in Gaussian-type limits. Intuitively, the covariance contributions from distant pairs are small enough that the variance of aggregated sums grows proportionally to the sample size rather than at a superlinear rate.
4 Dependence Structures That Admit Summability
4.1 Weak dependence heuristics
Covariance summability aligns with the idea of weak dependence: observations far apart may not be independent, but their correlation is sufficiently small. “Weak” here is quantitative: the cumulative effect of dependence across all separations must remain finite in the sense dictated by the chosen summability criterion.
4.2 Mixing conditions leading to summability
Many dependence models are characterized by mixing, which typically implies that correlations between distant events shrink. Under suitable moment conditions and stationarity, mixing rates can be converted into decay bounds for covariances, and hence into summability of \(\gamma(k)\). The exact pathway depends on the mixing type and regularity of the variables (e.g., whether one can relate \(\mathrm{Cov}(X_0,X_k)\) to mixing coefficients and boundedness or integrability).
4.3 Decay rates of correlations and covariance series
| Summability can often be read off from how quickly correlations decay. If \( | \gamma(k) | \) behaves like \(k^{-\alpha}\) for large \(k\), then convergence of \(\sum_k | \gamma(k) | \) requires \(\alpha>1\) in one-dimensional indexing. Exponential decay typically implies absolute summability automatically. |
|---|
4.3.1 Polynomial vs exponential decay examples
| - Polynomial decay: If \(\gamma(k)\approx Ck^{-\alpha}\), then \(\sum_{k\ge1} | \gamma(k) | \) converges when \(\alpha>1\), and diverges when \(\alpha\le 1\). In the divergent regime, variances of partial sums may grow faster than linear. | ||
|---|---|---|---|---|
| - Exponential decay: If \( | \gamma(k) | \le Ce^{-ck}\), then \(\sum_{k\ge1} | \gamma(k) | \) converges for any \(c>0\), yielding a stable asymptotic variance. |
5 Connections to Limit Theorems
5.1 Central limit theorem heuristics for dependent data
For independent data, the central limit theorem (CLT) relies on the variance scaling linearly and on approximate “averaging out” of fluctuations. Under covariance summability, the variance of \(S_n\) grows like \(n\sigma^2\), and the long-run dependence does not overwhelm the aggregation. Many CLT results for weakly dependent sequences further require additional structure beyond covariance summability (such as functional conditions, martingale approximations, or stronger mixing/moment assumptions), but summability is a key ingredient that prevents variance blow-up.
5.2 Law of large numbers under covariance summability
A law of large numbers (LLN) for dependent variables often uses variance bounds for averages \(\bar X_n=n^{-1}S_n\). Since \[ \mathrm{Var}(\bar X_n)=\frac{1}{n^2}\mathrm{Var}(S_n), \] linear variance growth combined with summability yields \(\mathrm{Var}(\bar X_n)\to 0\), enabling convergence in mean square and, with additional arguments, almost sure or in-probability forms. Thus, covariance summability serves as a mechanism to enforce that average fluctuations vanish asymptotically.
5.3 Functional limit theorems and invariance principles
Beyond pointwise convergence, one may seek convergence of the partial-sum process \(\{n^{-1/2}S_{\lfloor nt\rfloor}\}_{t\in[0,1]}\) to a Brownian motion (or a related Gaussian process). Covariance summability helps ensure tightness and determines the scaling variance through \(\sigma^2\). However, functional invariance principles typically require further regularity to control increments over short intervals.
6 Spectral Viewpoint
6.1 Covariance summability and spectral densities
For a stationary process, the covariance function \(\gamma(k)\) can be interpreted through a spectral representation. When \(\gamma(k)\) is absolutely summable, its Fourier series defines a continuous spectral density: \[ f(\omega)=\frac{1}{2\pi}\sum_{k=-\infty}^{\infty}\gamma(k)e^{-ik\omega}. \] This creates a direct bridge between time-domain decay and frequency-domain regularity.
6.2 Fourier transform links for stationary processes
6.2.1 Existence and continuity of spectra
Absolute summability guarantees the Fourier series converges and the resulting \(f(\omega)\) is bounded and continuous. Without such summability, one can still define spectral measures in a more general sense, but the clean density-based description may fail or require weaker forms of convergence.
6.3 Interpretation of asymptotic variance through spectra
The asymptotic variance \(\sigma^2\) often equals the spectral density integrated at the zero frequency: \[ \sigma^2 = 2\pi f(0), \] when the density representation holds. This interpretation clarifies why long-run dependence at low frequencies inflates variance: strong persistence corresponds to spectral mass concentrated near \(\omega=0\).
7 Sufficient Conditions and Inequalities
7.1 Bounding covariances via moment conditions
| Summability statements can be strengthened or made operational using moment assumptions. For example, if the variables have finite moments and some form of weak dependence holds, one can bound \( | \mathrm{Cov}(X_0,X_k) | \) by functions of dependence strength and integrability. Such bounds then translate into convergence or summability of the covariance series. |
|---|
7.2 Use of Cauchy–Schwarz and covariance inequalities
Basic inequalities can provide general-purpose estimates: \[
| \mathrm{Cov}(X,Y) | \le \sqrt{\mathrm{Var}(X)\mathrm{Var}(Y)}. |
|---|
\] While this alone does not guarantee summability, it becomes useful when combined with decay of dependence indicators or with bounds on \(\mathrm{Var}(X_t)\) and on how \(\mathrm{Cov}(X_0,X_k)\) decreases with \(k\). Other covariance inequalities tailored to mixing processes can yield directly summable upper bounds.
7.3 Relation to ergodicity and regularity assumptions
Although covariance summability is not identical to ergodicity, both are connected through how dependence behaves over long time horizons. Regularity conditions (e.g., stationarity, existence of moments, and suitable mixing-type behavior) often accompany summability criteria in limit theorem proofs, ensuring that averages stabilize and that variance calculations reflect the long-run structure.
8 Practical Considerations and Estimation
8.1 Empirical covariance lag plots
In applications, one rarely knows \(\gamma(k)\) exactly. Practitioners compute sample covariances for a range of lags and inspect their decay. A common diagnostic is a lag plot of \(\hat\gamma(k)\) to gauge whether the magnitude appears to fall fast enough to plausibly support summability.
8.2 Truncation and windowing strategies
Because the covariance series is infinite in theory, estimation typically truncates at a maximum lag \(K_n\) and may apply weights: \[ \hat\sigma^2 = \hat\gamma(0) + 2\sum_{k=1}^{K_n} w(k)\hat\gamma(k). \] Windowing weights reduce variance of the estimator by downweighting noisy large-lag estimates, at the cost of introducing bias if the tail is not negligible.
8.3 Estimating asymptotic variance under summability assumptions
8.3.1 Bias–variance tradeoffs in truncated sums
Truncation error decreases as \(K_n\) increases, but sampling variability increases because estimates of \(\gamma(k)\) at large lags are noisier. Summability implies the tail contribution shrinks, making it feasible to choose \(K_n\) large enough for bias control while still keeping estimation variance manageable.
9 Examples and Worked Scenarios
9.1 Moving average processes
| For a finite-order moving average \(X_t=\sum_{j=0}^q a_j \varepsilon_{t-j}\) with white noise \(\varepsilon_t\), covariances vanish beyond lag \(q\). Consequently, \(\sum_k | \gamma(k) | \) is finite, and asymptotic variance exists and is easily computed from the coefficients. |
|---|
9.2 Autoregressive models with short-range dependence
In many autoregressive models where dependence decays geometrically (e.g., stable AR processes), covariances typically decrease exponentially with lag. This produces absolute summability of \(\gamma(k)\), ensuring linear variance growth for partial sums and enabling standard asymptotic variance formulas used in CLT and LLN settings.
9.3 Long-memory counterexamples (summability failure)
| In long-memory models, covariances decay too slowly, often resembling \(k^{-\alpha}\) with \(\alpha\le 1\). Then \(\sum_{k\ge1} | \gamma(k) | \) diverges, and variance of partial sums can grow faster than \(n\). As a result, normalization by \(n^{1/2}\) may be inappropriate, and classical CLT forms may fail or require different scaling and limiting processes. |
|---|
10 Common Pitfalls and Edge Cases
10.1 Nonstationarity and ill-defined covariances
Covariance summability is usually defined in a stationary framework where \(\gamma(k)\) depends only on lag. With nonstationary sequences, covariances may not be well-defined in a lag-only form, or their magnitude may change over time, complicating whether a summability test is meaningful.
10.2 Conditional heteroskedasticity vs covariance decay
Processes may exhibit volatility clustering: conditional variances vary over time even if mean covariances decay. Covariance summability concerns second-order cross-covariances of the centered variables, which may still be small, but in practice one must distinguish between decay of \(\mathrm{Cov}(X_t,X_{t+k})\) and dependence in higher-order moments or in conditional structures.
10.3 Divergent series despite covariance terms tending to zero
| It is possible for \(\gamma(k)\to 0\) while \(\sum_k | \gamma(k) | \) diverges. The failure comes from insufficient decay rate: many small covariances can still add up to an unbounded total. This is a central reason summability conditions emphasize series convergence rather than mere pointwise decay. |
|---|
11 Summary and Key Takeaways
11.1 How to check summability in practice
A practical workflow is to:
- Verify stationarity or use methods appropriate for approximate stationarity.
- Estimate covariance lags and examine whether they appear summable in magnitude.
- If relying on a model class, use known decay rates or theoretical bounds to justify summability.
- When estimating long-run variance, use truncation and windowing with attention to the choice of maximum lag.
11.2 Typical consequences for asymptotic behavior
When covariance summability holds, the variance of partial sums typically grows linearly with the sample size, enabling the definition of a finite asymptotic variance and supporting standard LLN and CLT-type conclusions. Spectrally, it corresponds to regular behavior of the spectral density near zero frequency, preventing low-frequency dependence from dominating the scaling of aggregated fluctuations.