Multiband Periodogram

• Multiband periodogram is a statistical tool that extracts periodic signals from irregularly sampled multiband time series by leveraging joint information across channels.
• It applies advanced methodologies like Tikhonov-regularized shared-phase models and multiband AoV to control overfitting and enhance period detection accuracy.
• This approach is vital for large-scale surveys and sensor arrays, boosting period recovery rates in sparse and heterogeneous observational data.

A multiband periodogram is a statistical tool for detecting and characterizing periodic signals in multichannel, multiband, or multi-filter time series, where each channel (or band) may be observed at different, non-uniformly sampled epochs and with heterogeneous photometric uncertainties. The multiband periodogram generalizes classic single-band methods such as the Lomb-Scargle periodogram and the Analysis of Variance (AoV) periodogram by exploiting the joint information content of all available bands, improving sensitivity for period detection especially in sparsely and unevenly sampled multiband astronomical data. Recent techniques leverage joint modeling across bands, flexible spectral decomposition, and principled regularization. The multiband periodogram is relevant for large-scale time-domain surveys, compressive spectral estimation, and source detection in sensor arrays (VanderPlas et al., 2015, Mondrik et al., 2015, Ariananda et al., 2014, Liu et al., 2019).

1. Multiband Periodogram Formulations

Two notable multiband periodogram approaches have been developed:

• Shared-phase, Tikhonov-regularized multiband extension of Lomb-Scargle: Each band’s time series $y_k(t)$ is represented by a truncated Fourier series sharing a common frequency $\omega$. The model coefficients are decomposed into a base model (shared across bands) and per-band residuals. The optimization is over all bands jointly, with Tikhonov regularization penalizing per-band parameters to drive most variability into the base model, reducing overfitting and improving period recovery (VanderPlas et al., 2015).
• Multiband Generalization of Analysis of Variance (AoV) Periodogram: Each band is independently modeled as a $K$-harmonic Fourier series at a common trial frequency, with test statistics summed across bands. The aggregate F-ratio periodogram is maximized at the most likely frequency, providing a joint maximum-likelihood period estimate. This method dramatically improves completeness for low-cadence, non-simultaneous multiband observations (Mondrik et al., 2015).

Both approaches preserve classic features of Lomb-Scargle/AoV periodograms—treatment of heteroscedastic errors, applicability to non-uniform sampling—while leveraging cross-band information for increased sensitivity and reliability.

2. Mathematical Foundations and Algorithmic Structure

Shared-phase Multiband Model

For $K$ bands, the light curve in band $k$ is modeled as

$$y_k(t) = \mathrm{base}(t) + \mathrm{resid}_k(t) + \varepsilon_k(t),$$

with

$$\mathrm{base}(t) = \theta_0 + \sum_{n=1}^{N_\mathrm{base}} [\theta_{2n-1} \sin(n\omega t) + \theta_{2n} \cos(n\omega t)],$$

and per-band residual

$$\mathrm{resid}_k(t) = \theta_0^{(k)} + \sum_{n=1}^{N_\mathrm{band}} [\theta_{2n-1}^{(k)} \sin(n\omega t) + \theta_{2n}^{(k)} \cos(n\omega t)].$$

The model parameters $\theta$ are estimated via Tikhonov-regularized (ridge) weighted least squares: $$\min_\theta (y - X\theta)^\top \Sigma^{-1} (y - X\theta) + \theta^\top \Lambda\theta,$$ where $\omega$0 is the diagonal error covariance, and $\omega$1 encodes the regularization penalties, with per-band columns typically regularized more strongly than the shared base. The periodogram power is then evaluated as the variance explained, normalized across all bands: $\omega$2

Multiband AoV

Each band $\omega$3 with $\omega$4 observations is modeled as

$\omega$5

with the weighted residual sum of squares (RSS) computed for each band as a function of trial frequency $\omega$6. The multiband AoV periodogram at $\omega$7 is

$\omega$8

where $\omega$9 is the null (constant model) RSS in band $K$0 (Mondrik et al., 2015).

Compressive and Averaged Periodograms

Extensions for compressive sampling, binning, and sensor arrays restructure the periodogram calculation to operate on groups of frequency/angle bins, leveraging circulant or Toeplitz structures to enable strong compression and efficient reconstruction under suitable sampling patterns (Ariananda et al., 2014, Liu et al., 2019).

3. Regularization, Complexity, and Practical Implementation

Tikhonov Regularization is essential to stabilize the solution in the shared-phase multiband model. By applying strong penalties ($K$1) to per-band residuals and little or no penalty to the shared base model, variability is forced into common coefficients unless statistically justified otherwise. A typical choice is to set per-band regularization strength as a small fraction (e.g., $K$2) of the trace of the Fisher information matrix $K$3. As $K$4, the corresponding parameter is unconstrained; as $K$5, it is suppressed (VanderPlas et al., 2015).

Algorithmic Complexity: For the shared-phase method, the principal computation is solving $K$6 linear systems at each of $K$7 trial frequencies, with $K$8 (number of model coefficients) typically small compared to data size. Sparse design matrices and efficient solvers yield scaling $K$9 to $K$0. For the multiband AoV, complexity is $K$1, scalable with parallelization or using optimized BLAS/LAPACK libraries (VanderPlas et al., 2015, Mondrik et al., 2015).

Implementation: The shared-phase approach is implemented in the Python package gatspy. Usage involves providing time, magnitude, error, and band arrays, specifying the number of Fourier terms in the base and residuals, and then invoking fit and periodogram methods over a trial period grid.

$K$3 (VanderPlas et al., 2015)

4. Applications and Empirical Performance

Sparse, Non-Simultaneous Multiband Surveys

For RR Lyrae light curves in SDSS Stripe 82 (five bands, median 56 obs/band over 10 years), the multiband periodogram demonstrates superior performance over single-band and “Supersmoother” approaches. For data artificially sparsified to match future survey conditions (e.g., LSST after 6 months, with only 12–15 epochs per band), recovery rates using the multiband approach reach 64–94% (top-1/top-5 true period recovered in the result set) compared to as low as 32–45% for best single-band methods (VanderPlas et al., 2015, Mondrik et al., 2015).

A summary of reported period-recovery rates:

Scenario	Single-Band Recovery	Multiband Recovery
Dense (Stripe 82)	89% (top-1, Supersmoother)	79% (top-1), 99% (top-5)
Sparse (Stripe 82)	32% (top-1), 45% (top-5)	64% (top-1), 94% (top-5)
LSST, $K$2, 6mo	Supersmoother outperformed	>60% (top-5)

(VanderPlas et al., 2015, Mondrik et al., 2015)

The improvement arises because phase coverage is maximized when bands sample at independent times. Survey strategies that “dither” filter scheduling yield more effective multiband period recovery, while simultaneous observations in all bands give marginal improvement over single-band approaches.

Compressive and Array Applications

In array processing settings, the “multiband periodogram” or its compressive variants enable periodogram reconstruction from highly compressed or undersampled data using uniform binning and circulant structures for spectral or angular domains. The recovery is asymptotically unbiased given uncorrelated bin assumptions, and key design constraints (e.g., minimal circular sparse ruler sampling) guarantee identifiability. This framework underpins periodogram-based source enumeration schemes under low snapshot or underdetermined (more sources than sensors) conditions (Ariananda et al., 2014, Liu et al., 2019).

5. Extensions: Binning, Compression, and Source Enumeration

Compressive Periodogram

The compressive periodogram divides the frequency range into uniform bins and exploits coset correlation matrices with circulant properties, enabling reconstruction with substantially fewer samples by applying minimal circular sparse ruler sampling patterns. Least-squares reconstruction of bin powers, together with circulant matrix diagonalization, yields an efficiently computable, asymptotically unbiased compressed periodogram, with closed-form variance under Gaussian noise (Ariananda et al., 2014).

Source Enumeration via Multiband Periodogram Averaging

For underdetermined wideband source enumeration, multiband periodogram averaging is used to reinforce true direction-of-arrival (DoA) peaks across frequency bands, while averaging out frequency-dependent grating lobes and sidelobes. The process includes construction of a lag-redundancy-averaged Hermitian Toeplitz covariance matrix from the averaged periodogram, followed by eigenvalue analysis using the MDLgap information criterion. This method provably never overestimates the source count and is asymptotically consistent in equal-power settings (Liu et al., 2019).

6. Observational Cadence and Survey Strategy Implications

The efficacy of the multiband periodogram is highly sensitive to inter-band observing cadence. Non-simultaneous, interleaved band observations maximize phase coverage and joint information, significantly boosting period recovery rates. Block scheduling (simultaneous exposures in all filters) provides only marginal gains over single-band periodograms and should be avoided when period determination is a primary objective. For large-scale surveys such as LSST and the Dark Energy Survey, early operations should specifically aim to distribute band observations in time to take advantage of multiband periodogram performance (VanderPlas et al., 2015, Mondrik et al., 2015).

7. Limitations and Further Developments

While the multiband periodogram framework substantially improves sensitivity in the joint multiband setting, its relative gain is most pronounced in the sparse, heterogeneously sampled, and low per-band cadence regime. In richly-sampled, simultaneous-observation scenarios, the improvement over single-band approaches diminishes. The approach depends on correct specification of the shared-frequency model, assumes the period is identical across bands, and may require tuning of regularization parameters for optimal empirical performance. Extensions to compressive spatial/temporal sampling further require rigorous satisfaction of the identifiability conditions (e.g., minimal circular sparse rulers), and special care when bin correlations are present (Ariananda et al., 2014). Continued development addresses extensions to non-sinusoidal variability, broader astrophysical variability classes, and integration with nonlinear template-fitting for precise characterization.

References:

• Periodograms for Multiband Astronomical Time Series (VanderPlas et al., 2015)
• A Multiband Generalization of the Analysis of Variance Period Estimation Algorithm (Mondrik et al., 2015)
• Compressive Periodogram Reconstruction Using Uniform Binning (Ariananda et al., 2014)
• Wideband Source Enumeration Using Sparse Array Periodogram Averaging (Liu et al., 2019)