🔬 optics

Bayesian inversion for single-shot spectral-encoded waveform reconstruction

This paper presents a Bayesian inversion framework that successfully reconstructs ultrafast waveforms from single-shot spectral-encoded measurements by jointly inferring the THz field and probe parameters, thereby overcoming phase loss and dispersion-induced distortions without requiring additional optical complexity.

Original authors: Minsoo Kang, Thomas C. Underwood

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Minsoo Kang, Thomas C. Underwood

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

The Big Problem: The "Blurry Photo" of Light

Imagine you are trying to take a photo of a hummingbird's wings moving incredibly fast. Your camera shutter is too slow, so instead of a sharp image, you get a blurry mess. In the world of ultrafast science, scientists face a similar problem. They want to see how light and electricity behave in "ultrafast" moments (trillionths of a second), but their electronic detectors are too slow to catch the action directly.

To solve this, scientists use a trick called Spectral Encoding.

The Analogy: Imagine you have a long, stretched-out rubber band (a laser pulse). You paint a different color on every inch of the rubber band. If you stretch the rubber band out over time, the color at the start represents the start of the event, and the color at the end represents the end.
The Result: By looking at the colors (the spectrum) all at once, you can reconstruct what happened over time in a single snapshot. This is great for capturing things that happen only once or change too fast to measure normally.

The Catch: The "Distorted Mirror"

However, there is a flaw in this method. The process of stretching the rubber band and reading the colors acts like a distorted mirror.

Loss of Phase: When you look at the colors, you can see how bright they are (intensity), but you lose the information about how the waves are shifting (phase). It's like seeing a shadow of an object but not knowing its 3D shape.
The "Dead Zones": Because of the way the light is stretched, certain frequencies (colors) get completely wiped out, creating "nulls" or dead zones in the data. It's like trying to listen to a song, but the radio signal cuts out completely at specific notes.
The Distortion: The signal you get back is warped. If the original wave was a smooth hill, the measurement might look like a jagged mountain with weird ringing noises.

Traditionally, to fix this, scientists had to build complex, expensive hardware (like adding extra mirrors or lasers) to try to capture the missing information.

The Solution: A "Smart Detective" (Bayesian Inversion)

This paper introduces a new software-based solution. Instead of building new hardware, the authors created a mathematical detective called a Bayesian Inversion framework.

Here is how this detective works:

The Suspect (The Unknown Waveform): The detective doesn't know what the original wave looked like.
The Clues (The Messy Data): The detective is given the distorted, blurry photo (the single-shot measurement) and knows exactly how the camera lens distorts images (the physics of the setup).
The Rules (The Prior): This is the detective's secret weapon. Instead of guessing randomly, the detective uses "rules of thumb" based on physics.
- Analogy: If you are trying to guess the shape of a hidden object behind a curtain, and you know it's a "smooth, wavy object," you won't guess it's a jagged square. The paper uses a Gaussian Process as this rule. It tells the computer: "The wave should be smooth, and we know exactly where the 'dead zones' (nulls) are, so let's fill in the gaps logically based on the surrounding data."
The Verdict: The detective combines the messy clues with the rules to calculate the most likely original shape of the wave.

What They Did in the Lab

The team tested this "detective" using Terahertz (THz) waves (a type of light used to see through materials).

They created six different scenarios by changing how much they "stretched" the laser (changing the "chirp" rate).
In some cases, the stretching was so extreme that the "dead zones" wiped out huge chunks of the signal.
They compared their "smart detective" reconstruction against a standard, slow, high-precision measurement (the "gold standard" reference).

The Results

The results were impressive:

Restoring the Lost: The software successfully reconstructed the original wave, filling in the "dead zones" where the signal was supposed to be missing.
Accuracy: The reconstructed wave matched the high-precision reference measurement almost perfectly, even in the most distorted cases.
No New Hardware: They achieved this without adding a single new lens, mirror, or laser to their setup. They just changed the math.

The Bottom Line

This paper shows that you don't always need to buy better equipment to see faster or clearer. Sometimes, you just need a smarter way to interpret the data you already have. By using a "smart detective" approach (Bayesian inversion) that understands the rules of physics, they can undo the distortions caused by the measurement process itself, recovering the true shape of ultrafast waves that were previously thought to be lost.

Technical Summary: Bayesian Inversion for Single-Shot Spectral-Encoded Waveform Reconstruction

Problem Statement
Spectral encoding enables single-shot measurements of ultrafast transients by mapping temporal information onto the spectrum of a chirped probe pulse. This technique is critical for capturing non-repetitive or irreversible phenomena, such as plasma dynamics and electron diffraction, which exceed the response limits of conventional electronic detectors. However, the method suffers from fundamental limitations:

Phase Loss: Measurements record only spectral intensity, causing the loss of phase information in the encoded signals.
Dispersion-Induced Distortion: The detection process introduces dispersion, creating a tradeoff between temporal resolution and the measurement window. This manifests as waveform distortions and "spectral null frequencies" where signal information is lost.
Reconstruction Challenges: Existing inversion methods, such as Tikhonov regularization or Truncated Singular Value Decomposition (TSVD), often require prior knowledge of the undistorted waveform or rely on abstract singular mode spaces that cannot be guided by physical measurement features. Furthermore, these methods are highly sensitive to noise near spectral nulls, limiting the recoverable bandwidth.
Hardware Complexity: Previous attempts to mitigate these issues (e.g., Phase-diversity electro-optic sampling or dual-echelon detection) require additional optical hardware, increasing system complexity and alignment uncertainty.

Methodology
The authors propose a Bayesian inversion framework to recover the underlying terahertz (THz) waveform from the squared spectral observable without adding optical hardware. The approach treats the reconstruction as a probabilistic inference problem:

Forward Model: The framework utilizes a physical forward model describing the spectral encoding process. The measured intensity is modeled as the squared magnitude of the convolution between the chirped probe field and the target THz field, modulated by a kernel. This accounts for the non-linear distortion terms that arise from the finite temporal neighborhood of spectral sampling.
Parameterization: The inversion jointly infers:
- The target THz electric field, $E_{THz}(t)$ .
- Nuisance parameters, including the modulation coefficient $k$ and low-dimensional empirical parameters describing the probe spectrum.
Probe Characterization: The chirped probe field is characterized empirically using shot-to-shot measurements of the unmodulated probe spectrum and its group-delay profile. Principal Component Analysis (PCA) is applied to compress the probe ensemble, retaining leading components that capture dominant variations while modeling shot-to-shot deviations as latent weights with zero-mean Gaussian priors.
Priors:
- Gaussian Process (GP) Prior: A GP prior is assigned to the THz waveform. Crucially, the correlation scale of the GP is determined directly from the spectral null structure observed in the single-shot measurement. This allows the prior to span the null frequencies without requiring a separate delay-line reference measurement.
- Modulation Coefficient: A super-Gaussian prior is applied to the modulation factor $k$ , constrained within a physically estimated range.
Inference Algorithms: The posterior distribution is explored using two methods:
- Maximum A Posteriori (MAP): A quasi-Newton BFGS optimizer minimizes the negative log-posterior to find the most probable waveform.
- Metropolis-adjusted Langevin Algorithm (MALA): Used as a full-sampling benchmark to characterize the posterior distribution and uncertainty, preconditioned by the MAP Hessian.

Key Results
The framework was validated using single-shot THz time-domain spectroscopy (THz-TDS) experiments across six configurations, varying two probe spectral profiles (with and without a sapphire element) and three chirp conditions (using SF11 glass rods of 50, 100, and 150 mm, yielding chirp parameters $\alpha$ from 14.5 to 40 ps $^{-2}$ ).

Waveform Recovery: In all cases, the Bayesian inversion successfully reconstructed the time-domain THz waveform within the 95% credible interval of a delay-line reference measurement. The MAP estimates significantly reduced the residuals compared to raw single-shot data, particularly in regions of rapid field variation where distortion is most severe.
Spectral Null Recovery: The method effectively recovered spectral content lost at null frequencies. While raw single-shot spectra exhibited amplitude nulls and phase jumps, the MAP estimates filled these gaps, recovering the amplitude envelope and phase profiles consistent with the delay-line reference.
Robustness to Noise: The GP prior demonstrated superior noise robustness compared to SVD-based regularization. By retaining weak modes supported by the prior rather than discarding them based solely on singular value magnitude, the framework maintained a signal-to-noise ratio near unity at null frequencies, whereas Tikhonov regularization showed significant amplitude suppression.
Algorithmic Efficiency: The MAP estimate was found to be sufficient for experimental reconstruction, showing close agreement with the more computationally intensive MALA sampling, thereby enabling efficient processing of experimental data.

Significance and Claims
The paper claims to demonstrate the first algorithmic inversion of single-shot spectral-encoded waveforms that restores temporal resolution and recovers lost spectral content without the need for additional optical hardware or complex alignment.

Hardware Independence: The method eliminates penalties associated with the detection process by correcting distortions numerically, avoiding the need for extra optics (e.g., nonlinear crystals, delay lines) required by previous hardware-based solutions.
Reference-Free Operation: The framework reduces the dependency on a reference measurement for quantitative reconstruction. By deriving the prior directly from the probe beam characteristics and the spectral nulls of the measurement itself, it enables reconstruction even when the true target waveform is unknown a priori.
General Applicability: The authors posit that the explicit modeling of the forward problem and the use of physically informed priors provide a general pathway for quantitative single-shot spectroscopy. The formulation is adaptable to other spectral encoding modalities (e.g., pump-probe spectroscopy, photonic time-stretch) that suffer from similar phase-loss distortions, allowing existing spectrally encoded data to be reinterpreted at the software level.