---
title: 'MLE of Reconstructions: Methods & Trends'
url: https://www.emergentmind.com/topics/maximum-likelihood-estimation-of-reconstructions
type: topic
---

# MLE of Reconstructions: Methods & Trends

Maximum likelihood estimation of reconstructions is the practice of recovering a latent object from measurements by maximizing a likelihood, or equivalently minimizing a negative log-likelihood, under a specified forward model and noise model, and then using the recovered object as a plug-in surrogate for the unknown truth. In the literature, the reconstructed object ranges from a quantum state \(\rho\) to a Faraday spectrum represented by point sources, a nonnegative image \(x\), a ptychographic object or probe parameter vector \(\theta\), a speckle-free reflectivity image \(X\), or a motion-compensated PET activity measure \(\mu\) [1001.2029] [1211.5105] [2308.02436] [2602.10344] [2004.02444].

## 1. Statistical formulation

In its generic form, maximum likelihood estimation starts with data \(D\), an unknown parameter \(\theta\), and a statistical model \(p(D\mid \theta)\). The likelihood is \(L(\theta)=p(D\mid \theta)\), and the estimator is
\[
\hat\theta_{\mathrm{MLE}}=\arg\max_\theta L(\theta).
\]
The reconstructed object is then “plugged in” for prediction, compression, decision-making, or subsequent inference [1001.2029].

Across reconstruction problems, this template is specialized by the choice of forward operator and observation model. In quantum tomography with POVM elements \(\{\Pi_k\}\), the normalized negative log-likelihood is
\[
F(\rho)=-\sum_{k=1}^K f_k\log\bigl(\operatorname{tr}(\rho\Pi_k)\bigr),
\]
optimized over \(\rho\ge 0\) and \(\operatorname{tr}(\rho)=1\) [1609.07881]. In Poisson inverse problems, the negative log-likelihood becomes
\[
F(x)=\mathrm{KL}(y,Ax),
\]
with \(y\sim\mathrm{Poisson}(Ax)\) and \(x\in\mathbb{R}^n_{>0}\) [2602.13063]. In RM synthesis, assuming independent complex Gaussian noise across frequency channels, the negative log-likelihood is
\[
H(d\mid M)=\frac{1}{2}\sum_j\left[\frac{|d_j-\tilde d_j|^2}{\sigma_j^2}+\ln(2\pi\sigma_j^2)\right],
\]
where \(M=\{(m_i,\phi_i)\}\) is a continuous parametric model of point sources in Faraday depth [1211.5105]. In coherent imaging with mixed Poisson–Gaussian noise, the data term becomes
\[
L_{\mathrm{Mixed}}(\theta)=\sum_{k=1}^N\left[\ln\bigl(I_k(\theta)+\sigma_k^2\bigr)+\frac{(X_k-I_k(\theta))^2}{I_k(\theta)+\sigma_k^2}\right],
\]
which explicitly models signal-dependent shot noise and additive readout noise [2308.02436].

A recurring structural feature is that the maximum likelihood problem is rarely just a statistical fit. It is simultaneously a constrained inverse problem: the admissible parameter must satisfy physical, geometric, or structural constraints such as positivity, trace-one normalization, simplex constraints, sparsity, or a low-complexity tensor-network form [1609.07881] [2602.13063] [1308.2395].

## 2. Reconstruction objects and forward models

The reconstruction variable depends on the application domain, but the underlying pattern remains a physics-based map from object space to measurement space.

| Setting | Reconstruction variable | Likelihood or objective |
|---|---|---|
| Quantum state tomography | Density matrix \(\rho\) | \(L(\rho)=\prod_i[\operatorname{Tr}(\rho E_i)]^{n_i}\) |
| RM synthesis imaging | \(M=\{(m_i,\phi_i)\}_{i=1}^{N_M}\) | \(H(d\mid M)=\frac12\sum_j[\frac{|d_j-\tilde d_j|^2}{\sigma_j^2}+\ln(2\pi\sigma_j^2)]\) |
| Poisson image reconstruction | \(x\in\mathbb{R}^n_+\) | \(F(x)=\mathrm{KL}(y,Ax)\) |
| Ptychography | \(\theta\) with predicted intensities \(I_k(\theta)\) | \(L_{\mathrm{Mixed}}(\theta)\) for Poisson–Gaussian noise |
| Digital holography with speckle | \(x=\mathrm{vec}(X)\in(\mathbb{R}^+)^n\) | \(f_L(x)=\log\det\Sigma(x)+\frac1L\sum_\ell y_\ell^H\Sigma(x)^{-1}y_\ell\) |
| PET with continuous motion | \(\mu\in\mathcal{M}_+(K)\) | \(\ell(\mu)=\langle\mu,f\rangle-\sum_{\gamma\in\Gamma}\log\langle\mu,\gamma\rangle\) |

These examples illustrate that “reconstruction” may mean estimating a probability distribution, a measure, a list of off-grid point sources, or a high-dimensional field. In PET with continuous motion, for example, the object is a nonnegative Radon measure \(\mu\) on \(K\), motion is encoded by a time-dependent operator \(\mathcal{W}_t\), and list-mode detections are modeled by time-space Poisson point processes [2004.02444]. In digital holography with speckle, the likelihood is Gaussian in measurement space but with covariance
\[
\Sigma(x)=AXA^H+\sigma_z^2I,
\]
so the unknown reflectivity appears inside a matrix-valued covariance rather than only through a linear mean [2602.10344]. In algebraic statistics for Gaussian models, the reconstruction variable may even be a concentration matrix \(K\) constrained to lie in a 3-dimensional linear space \(L\subset S_3\), with log-likelihood
\[
\ell(K)=\log\det K-\operatorname{trace}(SK),
\]
turning reconstruction into an explicitly geometric problem on reciprocal surfaces of nets of conics [2011.08989].

This suggests a useful unifying view: maximum likelihood estimation of reconstructions is less about a single algorithm than about a class of estimation problems in which a latent object is defined indirectly through a measurement operator, and the statistical model is imposed in data space rather than object space.

## 3. Optimization strategies and computational scalability

Because exact closed-form solutions are rare, the quality of a reconstruction is inseparable from the optimization method used to obtain it. Several distinct algorithmic families appear repeatedly.

Expectation–maximization and its variants dominate Poisson models. Classical EMML for \(y\sim\mathrm{Poisson}(Ax)\) uses the multiplicative update
\[
x^{(k+1)}=x^{(k)}\odot \frac{A^\top \frac{y}{Ax^{(k)}}}{A^\top \mathbf{1}},
\]
preserving positivity by construction [2602.13063]. The same paper reinterprets EMML as mirror descent with mirror map \(\mathcal{A}(\theta)=\sum_{i,j}A_{ij}e^{\theta_j}\), which makes it possible to add convex constraints by Bregman projections while keeping the multiplicative structure. A related line of work treats EM as a Kullback proximal point algorithm,
\[
\theta^{k+1}=\arg\max_\theta \{l_y(\theta)-\beta_k I_y(\theta^k,\theta)\},
\]
recovering standard EM when \(\beta_k=1\), preserving monotonicity for \(\beta_k>0\), and obtaining superlinear convergence when \(\beta_k\to 0\) [1201.5907].

Projected-gradient methods are central in quantum tomography. In density-matrix space, the objective is convex and the feasible set is the quantum state space, so accelerated projected-gradient methods can operate directly on \(\rho\) rather than on factored parameterizations. In one benchmark, an 8-qubit product-Pauli MLE that reportedly took weeks in 2005 “now takes less than a minute” on a standard desktop using APG or CG-APG, while older direct-gradient and factored conjugate-gradient methods stalled far above the true maximum likelihood [1609.07881]. For many-body tomography, the same maximum-likelihood principle can be merged with matrix product state/operator representations, turning the update into an \(R\rho R\)-type iteration followed by tensor-network compression [1308.2395].

Other settings use maximum likelihood as a refinement stage on top of a heuristic initializer. In RM synthesis imaging, CLEAN or RMCLEAN first constructs a sparse but grid-locked model, and then maximum likelihood refines the continuous Faraday depths \(\phi_i\), complex amplitudes \(m_i\), and model order through BIC-guided component merging [1211.5105]. In digital holography with speckle, direct likelihood optimization is obstructed by the cost of repeated inversions of \(\Sigma(x)\). PGD-MC avoids explicit matrix inversions by combining conjugate gradient solves for \(\Sigma(x)^{-1}b\) with Monte Carlo diagonal estimation for \(\operatorname{diag}(A^H\Sigma(x)^{-1}A)\), making full-resolution MLE practical even with finite apertures [2602.10344].

A common misconception is that MLE is intrinsically computationally slow. The literature instead shows that apparent slowness is often algorithmic: it arises from suboptimal parameterizations, poor handling of constraints, or unnecessary approximations, rather than from the likelihood principle itself [1609.07881].

## 4. Regularization, constraints, and predictive robustness

Maximum likelihood reconstructions are often statistically admissible yet operationally brittle. The clearest example is quantum state tomography. Standard MLE frequently produces rank-deficient density matrices with zero eigenvalues, which can assign probability zero to outcomes that were merely unmeasured, not impossible. For predictive tasks such as gambling, coding, or future measurement prediction, this can lead to catastrophic behavior because \(D(\rho_{\mathrm{true}}\Vert\hat\rho)\) becomes infinite whenever the estimated support fails to contain the true support [1001.2029].

Hedged maximum likelihood estimation addresses this by modifying the objective with
\[
h(\rho)=\det(\rho)^\beta,\qquad
\ell'(\rho)=\ell(\rho)+\beta\log\det(\rho).
\]
The resulting estimator is strictly positive, reproduces the classical add-\(\beta\) rule in the single-basis limit, and satisfies
\[
\frac{L(\hat\rho_{\mathrm{HMLE}})}{L(\hat\rho_{\mathrm{MLE}})}\ge e^{-d\beta},
\]
so the gain in predictive robustness is obtained at only a small likelihood cost [1001.2029].

Constraints can also be incorporated directly rather than through penalties. Constrained EMML alternates the standard multiplicative Poisson step with a Bregman projection
\[
x^{(r+1)}=\operatorname{Proj}_C^u(\tilde x^{(r)})
=\arg\min_{x\in C} D_u(x,\tilde x^{(r)}),
\]
allowing simplex and related convex constraints to be enforced while preserving the EMML structure [2602.13063]. In unsupervised imaging, convex regularizers can themselves be learned by maximum marginal likelihood: the prior is parameterized as
\[
p_\theta(x)\propto e^{-g_\theta(x)},
\]
the MAP reconstruction is
\[
x_{\mathrm{MAP}}(y;\theta)=\arg\min_x\bigl(\ell(y\mid x)+g_\theta(x)\bigr),
\]
and the prior parameters are trained from noisy measurements by maximizing \(\log p(y\mid \theta)\), using the identity
\[
\nabla_\theta \log p(y\mid \theta)
=
\mathbb{E}_{x\sim p(x\mid \theta)}[\nabla_\theta g_\theta(x)]
-
\mathbb{E}_{x\sim p(x\mid y,\theta)}[\nabla_\theta g_\theta(x)].
\]
This links empirical Bayes training directly to classical variational reconstruction [2404.05445].

Another persistent concern has been the nonconvexity introduced by physical parameterizations. In quantum tomography, parameterizing \(\rho\) as \(T^\dagger T/\operatorname{tr}(T^\dagger T)\) creates a nonconvex unconstrained problem in the parameters \(t\), but when the negative log-likelihood is convex in \(\rho\), all local minima of the parameterized problem are global and correspond to the same density matrix. Practical failures are therefore attributed to numerical issues, not to “wrong” local minima [1103.3682].

## 5. Approximated, simulated, and marginal likelihoods

In some reconstruction problems the likelihood is itself unavailable in closed form. The estimation target then becomes a reconstructed likelihood rather than a directly evaluable one.

Approximate maximum likelihood estimation uses simulation when only a generative simulator is available. Given observed summary statistics \(S_{\mathrm{obs}}\), the likelihood \(p(S_{\mathrm{obs}}\mid \theta)\) is estimated by simulation and kernel density estimation, and the parameter is updated by stochastic approximation using simultaneous perturbation stochastic approximation. The estimator is therefore obtained by climbing a locally reconstructed likelihood surface rather than an analytic one [1507.04553].

Maximum approximated likelihood generalizes this perspective to deterministic numerical integration. If the exact likelihood contribution is an integral
\[
f(z,\theta)=\int_\Omega \varphi(v,z,\theta)\,\omega(v)\,dv
\]
and the approximation \(\tilde f_r(z,\theta)\) is obtained by Monte Carlo, quasi-Monte Carlo, Gaussian quadrature, or sparse-grid cubature, then the estimator maximizes
\[
\tilde L_{n,r}(\theta)=\frac1n\sum_{i=1}^n \log \tilde f_r(Z_i,\theta).
\]
Consistency follows when \(\sup_{z,\theta}|\tilde f_r-f|\to 0\), and asymptotic normality is recovered when the combined approximation error \(\mathcal{E}(R(n))\) satisfies \(\sqrt n\,\mathcal{E}(R(n))\to 0\) [1908.04110]. This provides a formal answer to when MLE on a numerically reconstructed likelihood still behaves like exact ML.

A more specialized use of the same idea appears in flow matching. Standard flow matching trains a vector field without integrating its ODE during training, creating a train–inference gap. Fine-tuning via maximum likelihood estimation of reconstructions assumes
\[
p_1(x\mid x_1)=\mathcal N(x\mid x_1,\Sigma),
\]
applies this likelihood to the terminal ODE output \(\hat\phi_N(x_0)\), and shows that for \(\Sigma=\sigma I\) maximizing the reconstruction likelihood is equivalent to minimizing
\[
\mathcal L_{\mathrm{MLE}}
=
\mathbb E_{q(x_0,x_1)}\|x_1-\hat\phi_N(x_0)\|^2.
\]
The paper proposes both straightforward fine-tuning and residual-based fine-tuning, the latter designed to incorporate contraction properties for robustness and interpretability [2510.02081].

These developments enlarge the notion of “maximum likelihood estimation of reconstructions.” It can refer not only to reconstructing an object by ML, but also to reconstructing the likelihood itself, or to fine-tuning a generative dynamical system by maximizing the likelihood of its reconstructions.

## 6. Empirical regimes, misconceptions, and limitations

The empirical record is heterogeneous but patterned. In quantum tomography, hedged MLE improves relative entropy, Euclidean distance, and infidelity for most single-qubit states, but “for nearly pure states, MLE does slightly better” and “neither method is optimal” [1001.2029]. In RM synthesis, maximum-likelihood refinement eliminates the dependence on pixelization that plagues CLEAN-based reconstructions; in the two-source tests, average errors under RMCLEAN are reported as “2+ orders of magnitude larger than ML” [1211.5105]. In ptychography, the mixed Poisson–Gaussian loss is reported to outperform the conventional Poisson-only loss under low signal-to-noise conditions, while the methods become nearly indistinguishable at high photon counts [2308.02436]. In continuous-motion PET, the derived ML-EM algorithm is monotone, convergent in the low-noise regime, and recovers the blur that would appear if movement were ignored [2004.02444].

Several misconceptions recur across fields. One is that making the discretization finer is sufficient; the RM synthesis results explicitly show that “even over-resolving the grid does not cure systematic errors” for multiple sources [1211.5105]. Another is that Gaussian approximations are harmless in high-dimensional quantum tomography; in the 8-qubit example, “82% of the outcomes had zero counts,” making Gaussian-noise methods “demonstrably inapplicable” in that regime [1609.07881]. A third is that plain maximum likelihood is automatically the best predictive estimator; the quantum hedging results show that a slightly less likely full-rank estimator can be much better for predictive tasks [1001.2029].

The principal limitations are equally consistent. Exact likelihoods may be expensive or unavailable; constraints may be hard to encode; and structurally correct models can still be computationally prohibitive without specialized optimization. Point-source assumptions restrict the RM-synthesis method when diffuse Faraday structure is present [1211.5105]. Convex-prior training by marginal likelihood is heavier than supervised training because it requires dual MCMC chains [2404.05445]. MLE in high-dimensional coherent imaging remains challenging because covariance-dependent likelihoods require repeated linear solves or trace surrogates [2602.10344].

Taken together, these works portray maximum likelihood estimation of reconstructions as a broad methodological family rather than a single doctrine. Its central strength is the explicit coupling of a physics-based forward model with a statistically specified data term. Its central weakness is that naive implementations can be overconfident, grid-locked, or computationally intractable. The mature forms of the method therefore combine likelihood principles with structural representations, constrained optimization, regularization, or controlled approximation, while remaining anchored in the measurement model itself [1609.07881] [1211.5105] [1001.2029].

Source: https://www.emergentmind.com/topics/maximum-likelihood-estimation-of-reconstructions