Uncertainty quantification for inverse problems with application to ptychographic reconstruction
Di, Zichao Wendy, Cherukara, Matthew J., Zhou, Tao · 2026
Read Uncertainty quantification for inverse problems with application to ptychographic reconstruction by Di, Zichao Wendy, Cherukara, Matthew J., Zhou, Tao (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/doe-osti-3020809-uncertainty-quantification-for-inverse-p.txt.
Introduction
Argonne National Laboratory report ANL-26/8, by Zichao Wendy Di, Matthew J. Cherukara, and Tao Zhou, works on a gap in scientific imaging: a reconstructed image is only half an answer without knowing how wrong it might be. The report develops uncertainty quantification for inverse problems, the class of problems where measurements must be turned back into the object that produced them. Ptychography, an X-ray imaging method used at synchrotrons, serves as the test application.
The approach builds on maximum a posteriori estimation, where the best reconstruction balances fit to data against prior expectations. The authors show how the curvature of that objective at its minimum, captured by its Hessian matrix, approximates the covariance of the estimate. They handle the awkward cases too: singular curvature calls for a pseudo-inverse, and non-smooth priors need smoothing or generalized Hessians. A preprint version is dated February 21, 2026.
Catalog record
| CREATOR | Di, Zichao Wendy, Cherukara, Matthew J., Zhou, Tao |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 8 KB |
| LICENSE | Public Domain |
| RIGHTS | PUBLIC DOMAIN |
| SOURCE | DOE OSTI |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
ANL-26/8
UNCERTAINTY
QUANTIFICATION FOR
INVERSE PROBLEMS
WITH APPLICATION TO
PTYCHOGRAPHIC
RECONSTRUCTION
Mathematics and Computer Science Division
About Argonne National Laboratory
Argonne is a U.S. Department of Energy laboratory managed by UChicago Argonne, LLC under
contract DE-AC02-06CH11357. The Laboratory’s main facility is outside Chicago, at 9700 South
Cass Avenue, Lemont, Illinois 60439. For information about Argonne and its pioneering science
and technology programs, see www.anl.gov.
DOCUMENT AVAILABILITY
Online Access: U.S. Department of Energy (DOE) reports produced after 1991 and a growing
number of pre-1991 documents are available free at OSTI.GOV (http://www.osti.gov/), a service
of the US Dept. of Energy’s Office of Scientific and Technical Information.
Reports not in digital format may be purchased by the public
from the National Technical Information Service (NTIS):
U.S. Department of Commerce
National Technical Information Service
5301 Shawnee Road
Alexandria, VA 22312
www.ntis.gov
Phone: 800-553-NTIS (6847) or 703-605-6000
Fax: 703-605-6900
Email: [email protected]
Reports not in digital format are available to DOE and DOE contractors
from the Office of Scientific and Technical Information (OSTI):
U.S. Department of Energy
Office of Scientific and Technical Information
P.O. Box 62
Oak Ridge, TN 37831-0062
www.osti.gov
Phone: 865-576-8401
Fax: 865-576-5728
Email: [email protected]
ANL-26/8
UNCERTAINTY
QUANTIFICATION FOR
INVERSE PROBLEMS
WITH APPLICATION TO
PTYCHOGRAPHIC
RECONSTRUCTION
by
Zichao Wendy Di (MCS/XSD, ANL), Mathew Cherukara (XSD, ANL)
FEBURARY 2026
U NCERTAINTY Q UANTIFICATION FOR I NVERSE P ROBLEMS WITH
A PPLICATION TO P TYCHOGRAPHIC R ECONSTRUCTION
A P REPRINT
Zichao Wendy Di
Mathematics and Computer Science Division & X-ray Science Division, Argonne National Laboratory
Mathew Cherukara
X-ray Science Division, Argonne National Laboratory
Tao Zhou
Center for Nanoscale Materials, Argonne National Laboratory
A BSTRACT
Inverse problems in imaging are commonly solved by optimization or learned surrogates that return
a single reconstruction, while uncertainty information is often unavailable. In many experimental
settings, however, uncertainty is required to assess reliability, guide downstream analysis, and
prioritize additional measurements. In this note, we present a compact uncertainty-quantification
framework based on local objective curvature, and then specialize it to ptychographic reconstruction.
We further show how repeated reconstructions can be aggregated in a statistically principled way,
including a practical implementation path for PtychoNN.
1 Introduction
Inverse problems aim to recover unknown parameters from indirect, noisy, and typically ill-conditioned measurements.
Because these problems are often non-unique or weakly constrained, point estimates alone provide incomplete
information: two reconstructions with similar data fidelity may have substantially different reliability across spatial
regions.
This motivates uncertainty quantification (UQ) that is local, interpretable, and compatible with existing reconstruction
pipelines. A useful route is curvature-based UQ, where local second-order structure of the objective function approxi-
mates posterior spread around a maximum a posteriori (MAP) estimate. This approach is attractive because it directly
leverages quantities already computed in optimization.
Ptychography provides a concrete and important setting for such methods. The forward model is nonlinear, the
parameter dimension is high, and gauge ambiguities are intrinsic. Recent learned methods such as PtychoNN provide
fast inference, but practical deployment still benefits from confidence maps and uncertainty-aware fusion across repeated
reconstructions.
2 Method
2.1 General inverse-problem formulation
Let x denote unknown parameters and let data be modeled by likelihood p(data | x). We consider MAP estimation
through
x̂ = arg min Ψ(x), Ψ(x) = −ℓ(x) + R(x),
x
where ℓ(x) = log p(data | x) and R collects prior or regularization terms.
The observed Fisher information for the likelihood is
Ilike (x) = −∇2x ℓ(x),
A PREPRINT - F EBRUARY 21, 2026
while the local curvature of the MAP objective is
∇2x Ψ(x) = −∇2x ℓ(x) + ∇2x R(x).
Hence the Hessian of Ψ coincides with likelihood Fisher information only when R ≡ 0.
Around a local minimizer x̂ with nonsingular curvature,
rh i
−1 −1
Cov(x̂) ≈ ∇2x Ψ(x̂) , σi ≈ (∇2x Ψ(x̂)) .
ii
In singular cases, a pseudo-inverse or additional regularization is required. For non-smooth priors (e.g., TV), one uses
smoothing, generalized Hessians, or local quadratic approximations.
2.2 Specialization to ptychography
In ptychography, object O and probe P are estimated from diffraction counts {yjk }. For scan position j,
2
uj = P ⊙ (Sj O), λjk = |(Fuj )k | + β,
where Sj is a shift/cropping operator and β is background. Under a Poisson model,
X
D(O, P ) = [λjk − yjk log λjk ] ,
j,k
and reconstruction solves
Ψ(O, P ) = D(O, P ) + αO RO (O) + αP RP (P ).
To apply curvature-based UQ, we parameterize complex unknowns in real coordinates
T
θ = Re (O) , Im (O) , Re (P ) , Im (P ) ,
and evaluate local curvature at θ̂:
H = ∇2θ Ψ(θ̂) or H ≈ J T W J + ∇2θ R.
Then
Cov(θ̂) ≈ H −1 (or H † under gauge constraints),
and pixelwise uncertainty follows from diagonal entries.
2.3 Principled aggregation of repeated reconstructions
Repeated reconstructions are common in practice (different random seeds, checkpoints, scan subsets, or noise real-
izations). For run r = 1, . . . , R and pixel/channel q, let θ̄r,q be the run-level estimate, s2r,q its run-level uncertainty
estimate, and nr an effective sample size. Define
R R
X 1 X
N= nr , θ̄q = nr θ̄r,q .
r=1
N r=1
The pooled uncertainty is
PR 2
PR 2
r=1 (nr − 1)sr,q + r=1 nr (θ̄r,q − θ̄q )
s2q = .
N −1
This decomposes into within-run and between-run components, separating local data/model uncertainty from multi-run
variability induced by nonconvexity or acquisition heterogeneity.
When full covariance is available per run, precision pooling yields
R
!−1 R ! R
!−1
X X X
θ̂pool = C−1r C−1
r θ̂
(r)
, Cpool = C−1
r .
r=1 r=1 r=1
2
A PREPRINT - F EBRUARY 21, 2026
2.4 PtychoNN-oriented workflow
For PtychoNN, this aggregation framework can be integrated after inference with minimal architectural changes:
1. run inference multiple times (R runs) with controlled perturbations (seed/checkpoint/subset/stochastic infer-
ence);
2. convert each run from amplitude-phase outputs to complex form,
zr,q = Ar,q exp(iϕr,q );
3. gauge-align each run to a common reference (global phase, shift, and probe/object scaling);
4. estimate run-level uncertainty s2r,q using either stochastic forward passes or short physics-based MAP refine-
ment with local curvature;
5. pool reconstructions with the weighted mean and pooled-variance formulas above;
6. recover pooled amplitude and phase from complex pooled estimates,
R
1 X
z̄q = nr zr,q , Āq = |z̄q |, ϕ̄q = arg(z̄q );
N r=1
7. export pooled estimate, pooled standard deviation, and within/between-run uncertainty maps for quality control
and adaptive reacquisition.
3
www.anl.gov
ARGONNE NATIONAL LABORATORY
¨ U.S. Department of Energy research facility
¨ Operated by the UChicago Argonne, LLC
¨ Midwest’s largest federally funded R&D facility
¨ Located in Lemont, IL, about 25 miles (40 km)
southwest of Chicago, IL (USA)
¨ Conducts basic and applied research
in dozens of fields
¨ Unique suite of leading-edge and rare scientific
user facilities
CONTACT
Argonne National Laboratory
9700 South Cass Avenue
Lemont, IL 60439