## Abstract

This document provides a correction to the derivation of the
distributions of the scattered fields presented in Optica **7**,
63 (2020) [CrossRef] and discusses the
implications this correction has on the rest of the paper.

© 2020 Optical Society of America under the terms of the OSA Open Access Publishing Agreement

## 1. SUMMARY OF CORRECTION

Our original derivation, presented in Section 3 of Ref. [1], modeled scattering off of the virtual source, the hidden object, and the virtual detector as modulating the incident field by a uniformly distributed random phasor and a constant equal to the square root of the albedo of the scattering surface. [Eq. (5) should have read $ {E_{{O_{\rm out}}}}({x_0}) = {E_{{O_{\rm in}}}}({x_O}){r^{1/2}}({x_O}) $].

In essence, this treats the scattering at position $ x $ as if the field is interacting with just a single scatterer. However, for any meaningful discretization, the field interacts with multiple dephased scatterers at each position $ x $ and the resulting scattered fields are summed together. Thus, the scattering process is more accurately modeled by multiplying the incident field by a complex random variable drawn from a circular complex Gaussian distribution [2].

In Section 2 we present the
corrected derivation of the field distributions. The main difference is
that the fields emitted from the hidden object and the virtual detector
follow circular complex *double* Gaussian
distributions [3]. That is, we have
speckled speckle. Additionally, we relax the far-field propagation
restriction by observing that the mathematical principles underlying
correlography can be extended to accommodate Fresnel propagation [4].

In Section 3 we discuss the implications these changes have on our results.

## 2. CORRECTED DERIVATION OF FIELDS

We assume that the virtual source surface is illuminated by a collimated beam at normal incidence so that

The incident field interacts with multiple scatterers on the optically rough virtual source. This causes the scattered field to follow a circular complex Gaussian distribution,

where $ {E_{V{S_{\rm out}}}}({x_{VS,1}}) $ and $ {E_{V{S_{\rm out}}}}({x_{VS,2}}) $ are independent for $ {x_{VS,1}} \ne {x_{VS,2}} $.The field emerging from the virtual source then undergoes free-space propagation on its way to the object, which can be modeled by a Fresnel transformation. Accordingly, the field incident on the hidden object is

where $ {\cal F} $ denotes the Fourier transform operator and $ {p_1} $ and $ {p_2} $ are quadratic phase terms with $ |{p_1}| = |{p_2}| = 1 $.From the Central Limit Theorem and the independence of the fields at different locations of $ {E_{V{S_{\rm out}}}} $ we have that for all $ {x_O} $, $ {\cal F}({p_1}{E_{V{S_{\rm out}}}})({x_O}) $ follows a circular complex Gaussian distribution. Because the phase of $ {\cal F}({p_1}{E_{V{S_{\rm out}}}})({x_O}) $ is uniformly distributed, multiplying it by a quadratic phase $ {p_2} $ does not change its distribution, thus $ {E_{{O_{\rm in}}}}({x_O}) $ also follows a circular complex Gaussian distribution.

Following the Van Cittert–Zernike theorem, the autocorrelation of $ {E_{{O_{\rm in}}}} $ is determined by the Fourier transform of the intensity of $ {E_{V{S_{\rm in}}}} $ [2]. Thus, assuming we illuminate a sufficiently large spot size on the virtual source, the autocorrelation of $ {E_{{O_{\rm in}}}} $ is $ \sigma _{{O_{\rm in}}}^2\delta (\Delta {x_O}) $ for some constant $ \sigma _{{O_{\rm in}}}^2 $.

At each location $ {x_O} $ on the hidden object, the incident field interacts with multiple scatterers. The resulting scattered field is

where for all $ {x_0} $, $ {g_O}({x_O}) $ follows an independent circular complex Gaussian distribution with variance $ r({x_O}) $, where $ r $ denotes the albedo of the hidden object at position $ {x_O} $.Accordingly, $ {E_{{O_{\rm out}}}} $ follows a circular complex double Gaussian distribution whose autocorrelation is given by

This field propagates to the virtual detector. Thus,

where $ {p_3} $ and $ {p_4} $ are quadratic phase terms. This incident field scatters of multiple scatterers on the virtual detector and we image the intensity of the scattered field. Thus,To make subsequent analysis simpler, we note that

where $ {\tilde E_{{O_{\rm out}}}} = {p_3}{E_{{O_{\rm out}}}} $. Because $ {E_{{O_{\rm out}}}} $ has uniformly distributed phase, $ {\tilde E_{{O_{\rm out}}}} $ and $ {E_{{O_{\rm out}}}} $ follow the same distribution.## 3. IMPLICATIONS

Correlography is based on relating the ensemble power spectral density (PSD) of the observed images to the autocorrelation function of the hidden object’s albedo. This derivation is based on two steps: (1) showing that the ensemble PSD converges to $ {\mathbb E}[|f \star f{|^2}|](\Delta x) $, where $ f $ is used to denote $ {\tilde E_{{O_{\rm out}}}} $. (2) showing that $ {\mathbb E}[|f \star f{|^2}|](\Delta x) $ is equal to the autocorrelation of the albedo, plus an additional term at $ \Delta x = 0 $.

Changing the distribution of $ {\tilde E_{{O_{\rm out}}}} $ and $ I $ requires us to slightly modify our derivation of each step. It also has minor implications on the predicted variance of our estimate.

#### A. PSD Estimate

Because $ |{\cal F}(f{)|^2} $ is modulated by speckle $ |{g_D}{|^2} $, as the number of observations goes to infinity our estimate of the autocorrelation converges to the PSD of $ |{g_D}{|^2}|{\cal F}(f{)|^2} $, rather than that of $ |{\cal F}(f{)|^2} $. We now relate the PSD of $ |{g_D}{|^2}|{\cal F}(f{)|^2} $ to the PSD of $ |{\cal F}(f{)|^2} $.

Let $ \beta = |{g_D}{|^2} $ and $ E = |{\cal F}(f{)|^2} $, with $ I = \beta E $. Assuming $ E $ is approximately wide sense stationary, $ I $’s autocorrelation function is

Thus the power spectral density of $ I $ is given by

where $ * $ denotes convolution.Because $ {R_\beta }(\Delta {x_D}) = {c_1} + {c_2}\delta (\Delta {x_D}) $ where $ {c_1} $ is a constant due to the non-zero mean of the speckle, the PSD of the speckle is given by $ {\cal F}({R_\beta })(\Delta x) = {c_2} + {c_1}\delta (\Delta x) $. Thus, the PSD of $ I $ is given by

Thus, when one accounts for speckle noise at the virtual detector, the ensemble PSD estimate converges to a scaled version of the unspeckled PSD (which we can equate to $ {\mathbb E}[|f \star f{|^2}|](\Delta x) $ using the original derivation found in the supplement of [1]) plus a constant offset term. This offset is already modeled by the $ b $ term in Eq. (11) of [1], though the paper assigns this term primarily to shot noise.

#### B. Autocorrelation Estimate

The next step of the proof shows that $ {\mathbb E}[|f \star f{|^2}|](\Delta x) $ is equal to the autocorrelation of the albedo plus a zero lag term. We repeat this proof here with the corrected scattering model. To make notation more concise, we let $ {z_1} = {p_3}{g_O} $ and $ {z_2} = {E_{{O_{\rm in}}}} $ and assume the limits of integration are from $ - \infty $ to $ \infty $ unless otherwise noted.

First, from the independence of $ {z_1} $ and $ {z_2} $ we have that

Using Isserlis’ theorem, substituting in the values of the covariances of $ {z_1} $ and $ {z_2} $, and assuming $ \sigma _{{O_{\rm in}}}^2 = 1 $ we have

As desired, $ {\mathbb E}[|(f \star f{)|^2}](\Delta x) $ is still equal to the autocorrelation of the albedo, plus an offset at $ \Delta x = 0 $. While this offset changes with the corrected scattering model, it plays no role in how we generate training data nor reconstruct the hidden object.

#### C. Variance Estimate

The paper argues that $ I $ follows a Gaussian-like distribution, and thus, following [5], the variance of the autocorrelation estimate is proportional to the true PSD squared divided by the number of observations. However, by incorporating the correct scattering model we find that, ignoring noise, as speckled speckle $ I $ follows a $ K $-distribution (page 60 of [2]), which is not well approximated by a Gaussian. While our empirical results suggest the aforementioned result remains valid, our argument is more heuristic than originally claimed.

## Funding

Defense Advanced Research Projects Agency (HR0011-16-C-0028).

## Disclosures

The authors declare that there are no conflicts of interest related to this paper.

## REFERENCES

**1. **C. A. Metzler, F. Heide, P. Rangarajan, M. M. Balaji, A. Viswanath, A. Veeraraghavan, and R. G. Baraniuk, “Deep-inverse correlography:
towards real-time high-resolution non-line-of-sight
imaging,” Optica **7**,
63–71 (2020). [CrossRef]

**2. **J. W. Goodman, *Speckle Phenomena in Optics: Theory
and Applications*, 2nd ed.
(Roberts and Company Publishers,
2020).

**3. **N. O’Donoughue and J. M. Moura, “On the product of independent
complex Gaussians,” IEEE Trans. Signal
Process. **60**,
1050–1063 (2011). [CrossRef]

**4. **J. R. Fienup, “Emerging systems and
technologies,” in *The Infrared and
Electro-Optical Systems Handbook*, Atmospheric
propagation of radiation (Infrared
Information Analysis Center, 1993),
Vol. 8, Chap. 1.5.

**5. **P. Welch, “The use of fast fourier
transform for the estimation of power spectra: a method based on time
averaging over short, modified periodograms,”
IEEE Trans. Audio Electroacoust. **15**,
70–73 (1967). [CrossRef]