Valley-Peak Modulation in Phase Space: an Exposure-Invariant VPM and its Theta-Function Structure

This paper introduces an exposure-invariant Valley-Peak Modulation (VPM) metric for deep sub-electron read noise sensors by mapping the Poisson-Gaussian model to a wrapped-Gaussian phase space, revealing that VPM follows a theta-function structure dependent solely on read noise and recovering existing approximations as truncated lattice sums.

The Gist

Imagine you are trying to listen to a very quiet conversation in a noisy room. The "conversation" is the signal from a camera sensor (how much light hit the pixel), and the "noise" is the static in the room (read noise).

In modern, super-sensitive cameras, the signal is so weak that it comes in tiny, distinct packets called photons (or electrons). If the camera is perfect, you'd see a perfect staircase: 1 photon here, 2 photons there, 3 photons there. But because of the "static" (noise), those steps get blurry. Instead of a sharp step, you get a fuzzy hill.

The Problem: The "Valley" is Hard to Measure

Scientists want to measure exactly how much "static" (noise) is in the camera. They look at the space between the hills (the "valleys").

• The Old Way: They tried to measure the depth of the valley between two hills.
• The Catch: The depth of the valley depends on two things:
  1. How much noise there is (the blurriness).
  2. How many photons are hitting the sensor (the exposure).

If you change the lighting (exposure), the hills get taller or shorter, and the valley looks different. This makes it hard to get a consistent measurement of the noise. It's like trying to judge how foggy a day is by looking at the distance between two trees, but the trees keep moving closer or further apart depending on the wind.

The Solution: The "Phase Space" Trick

The authors of this paper, Aaron and David, came up with a clever mathematical trick to solve this. They realized that if you ignore the total number of photons and only look at the remainder, the noise measurement becomes perfectly stable, no matter how bright or dark the scene is.

Here is the analogy:

The Clock Analogy
Imagine the number of photons is like the time on a clock.

• If you have 1 photon, it's 1:00.
• If you have 2 photons, it's 2:00.
• If you have 100 photons, it's 100:00.

The "noise" makes the hands of the clock jittery. The old method tried to measure the jitter by looking at the distance between 1:00 and 2:00. But if you have 100 photons, you are looking at 100:00 and 101:00, which changes the context.

The authors said: "Let's just look at the clock face itself."
They took the total number of photons and divided by 1, keeping only the remainder (the decimal part).

• 1.2 photons becomes 0.2.
• 2.2 photons becomes 0.2.
• 100.2 photons becomes 0.2.

By doing this, they "wrapped" all the data onto a single circle (like a clock face from 0 to 1). This is what they call Phase Space.

The "Wrapped Gaussian" (The Fuzzy Circle)

Once they wrapped the data onto this circle:

• The "hills" (peaks) and "valleys" of the data all line up perfectly on the circle, regardless of how many photons were originally there.
• The "noise" simply smears the data around the circle.
• If the noise is low, the data stays in a tight clump (a sharp peak).
• If the noise is high, the data spreads out evenly around the circle (a flat line).

Because they removed the "total count" variable, the shape of this smear only depends on the noise. It is now exposure-invariant. It's like measuring the fog by looking at how much a single candle's light spreads, rather than trying to measure the distance between two moving cars.

The "Theta Function" (The Secret Code)

The paper mentions something called a "Theta Function." In simple terms, this is just a fancy mathematical formula that perfectly describes the shape of that smeared circle.

• The authors found a "closed-form" (a neat, exact equation) for this shape.
• They also found a way to reverse the equation. If you measure the "smear" (the Valley-Peak Modulation), you can plug it into their formula and instantly calculate the exact amount of noise, without needing to guess or run complex simulations.

Why This Matters

1. Simplicity: It turns a messy problem (where lighting changes everything) into a clean, stable one.
2. Accuracy: It explains why previous "quick fixes" (approximations) worked well in low-light conditions—they were just simplified versions of this exact "wrapped circle" math.
3. Future Cameras: As cameras get better and better at seeing in the dark (Deep Sub-electron Read Noise), this method gives engineers a precise ruler to measure just how good their sensors really are, regardless of the lighting conditions.

In a nutshell: The authors realized that to measure the "fuzziness" of a camera sensor, you shouldn't look at the whole picture. Instead, you should wrap the data around a circle, ignore the total count, and measure how much the "fuzz" spreads around that circle. This gives you a perfect, unchanging ruler for camera noise.

Technical Summary

1. Problem Statement

In deep sub-electron read noise (DSERN) CMOS sensors, the quantization of photoelectron statistics becomes observable, allowing for characterization via photon counting histograms (PCH). A key metric for quantifying read noise ($\sigma$) in this regime is Valley-Peak Modulation (VPM), which measures the depth of the "valleys" between integer photoelectron peaks relative to the peak heights.

However, the original definition of VPM (in the amplitude domain) is a function of both read noise and the quanta exposure ($H$). While Starkey & Fossum previously derived exposure-independent approximations for small read noise, the underlying mathematical object that these approximations target remained undefined. The authors aim to identify the exact, exposure-invariant quantity that governs VPM and to provide a rigorous mathematical framework for its calculation and inversion.

2. Methodology

The authors propose a transformation from the amplitude domain to a phase space to eliminate the dependence on quanta exposure ($H$).

• Phase Mapping: Starting from the standard linear sensor model $X = \mu + \frac{1}{g}(K + \sigma Z)$, where $K$ is the integer photoelectron count and $Z$ is Gaussian noise, the authors apply a modulo-1 operation. They define a phase variable $\Phi$ using the unit-circle embedding:
  $$\Phi = \text{Arg}\left(e^{i 2\pi g(X-\mu)}\right)$$
  This operation effectively quotients out the integer photoelectron count ($K$), mapping the data onto the interval $(-\pi, \pi]$.
• Statistical Derivation:
  • The resulting variable $\Phi$ follows a Wrapped Gaussian distribution with a density independent of $H$.
  • Using Poisson summation, the probability density function (PDF) of $\Phi$ is expressed as a Jacobi theta function ($\vartheta_3$) with a nome $q = e^{-2(\pi\sigma)^2}$.
  • The PDF is given by: $f_\Phi(\phi) = \frac{1}{2\pi} \vartheta_3\left(\frac{\phi}{2}, q\right)$.
• Definition of Phase-Space VPM: The authors define the exposure-invariant VPM ($VPM_\phi$) as the ratio of the valley density (at $\pi$) to the peak density (at $0$):
  $$VPM_\phi(\sigma) = 1 - \frac{f_\Phi(\pi)}{f_\Phi(0)} = 1 - \frac{\vartheta_4(q)}{\vartheta_3(q)}$$
• Inverse Mapping: The authors derive a closed-form expression to recover read noise $\sigma$ from a measured $VPM_\phi$ using elliptic integrals. This involves inverting the relationship between the nome $q$ and the elliptic modulus $m$.

3. Key Contributions

1. Exposure-Invariant Formulation: The paper proves that the exposure dependence in standard VPM arises solely from the Poisson weighting of adjacent peaks. By mapping to phase space, the authors isolate a metric ($VPM_\phi$) that depends only on read noise ($\sigma$).
2. Theta-Function Representation: The authors provide an exact closed-form expression for VPM using Jacobi theta functions, replacing the need for numerical integration of the Poisson-Gaussian mixture.
3. Unification of Approximations: The paper demonstrates that the three exposure-independent approximations previously proposed by Starkey & Fossum are simply low-order truncations of the lattice-sum series expansion of the phase-space VPM. This explains their asymptotic accuracy in the DSERN regime.
4. Analytical Inversion: A novel, explicit formula is derived to calculate read noise directly from VPM measurements using complete elliptic integrals of the first kind ($K(m)$), avoiding iterative numerical methods.

4. Results

• Theoretical Validation: The authors show that as read noise ($\sigma$) approaches zero, the phase-space VPM asymptotically approaches 1 (indicating perfect valley separation), while for large $\sigma$, it approaches 0 (uniform distribution).
• Approximation Accuracy: Simulations confirm that the Starkey & Fossum formulas (Eq. 2a–2c) converge to the exact phase-space VPM as $\sigma \to 0$. The paper identifies $\sigma < 0.5$ e- as the regime where VPM becomes distinct, and $\sigma < 0.2$ e- as the ultra-low-noise regime where VPM $\approx 1$.
• Simulation Example: A simulation with $N = 5 \times 10^6$ observations and ground-truth parameters ($H=3, \sigma=0.2$) was performed.
  • The phase transform was applied to simulated gray values.
  • The estimated VPM from the phase histogram yielded a read noise estimate of $\hat{\sigma} = 0.2027$ e-.
  • This result is in close agreement with the true value of $0.2$ e-, validating the practical utility of the inverse mapping.

5. Significance

This work fundamentally reframes the characterization of read noise in DSERN sensors. By shifting the perspective from the amplitude domain to the phase domain, the authors:

• Decouple read noise estimation from exposure levels, simplifying calibration and characterization protocols.
• Provide a rigorous mathematical foundation for existing heuristic approximations, clarifying their limits and validity.
• Offer closed-form tools (theta functions and elliptic integrals) that enable faster, more precise, and analytically tractable read noise estimation without relying on heavy numerical simulations or exposure-specific calibration curves.

The paper concludes that while it does not propose a new end-to-end characterization method, it isolates the "invariant object" underlying VPM, providing essential theoretical tools for future sensor characterization research.