Projected Distribution Loss Overview

• Projected Distribution Loss is a concept with field-specific definitions, ranging from a perceptual training objective in deep learning to a forecasted loss distribution in risk and energy applications.
• In image enhancement and autoencoder tasks, it leverages projected feature distributions—often evaluated via Wasserstein distances—to balance reconstruction accuracy with perceptual quality.
• In actuarial and power system studies, the term describes modeled loss distributions or measurable technical losses, underpinning improved risk assessment and operational optimization.

to=python.exec code to=shell 彩票总代 彩神争霸大发快三 code to=shell 弘鼎 】【。】【”】【code to=shell 天天中彩票一等奖 code='["bash","-lc","python - <<"ERR\", e)\nPY"]' to=shell 彩神争霸如何 code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?search_query=all:%22Projected+Distribution+Loss%22&start=0&max_results=5'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(data[:1200])\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?search_query=all:%22Projected+Distribution+Loss%22&start=0&max_results=5'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(data[:1200])\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nids=['(Delbracio et al., 2020)','(Orioua et al., 12 Sep 2025)','(Mohlin et al., 2023)','(Avanzi et al., 2024)','(Fatima et al., 11 May 2026)']\nfor aid in ids:\n url=f'http://export.arxiv.org/api/query?id_list={aid}'\n try:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n title=''\n if '<title>' in data:\n parts=data.split('<entry>')\n if len(parts)>1:\n seg=parts[1]\n t=seg.split('<title>')[1].split('</title>')[0].strip().replace('\n',' ')\n title=t\n print(aid, title)\n except Exception as e:\n print('ERR', aid, e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?id_list=([2012.09289](/papers/2012.09289))'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(repr(data[:500]))\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","echo hi"]} “Projected Distribution Loss” is not a single standardized construct across arXiv literature. The expression and closely related variants are used for several technically distinct objects: a perceptual training objective for image enhancement based on projected feature distributions; a projected reconstruction term for intrinsic-dimension estimation in autoencoders; a camera-aware negative log-likelihood for probabilistic 3D regression; a forecast of the full conditional law of losses in actuarial modeling; a compound aggregate loss distribution for financial crises; and the estimation, reduction, or allocation of technical losses in electrical distribution networks (Delbracio et al., 2020, Orioua et al., 12 Sep 2025, Mohlin et al., 2023, Avanzi et al., 2024, Kapp et al., 2012, Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020). This suggests that the term is field-dependent: in some papers it denotes a training loss built from projected representations, while in others it denotes a projected or estimated distribution of losses, or physical loss in distribution systems.

1. Terminological scope and recurrent idea

Across the cited literature, the shared motif is the use of a projection, refinement, or conditional distribution to represent uncertainty or loss more faithfully than a simple point estimate. The technical object, however, changes by domain.

Domain	Meaning in the source literature	Source
Image enhancement	Perceptual loss using aggregated 1D-Wasserstein distances between projected CNN features	(Delbracio et al., 2020)
Intrinsic-dimension estimation	Projected reconstruction loss after removing one active latent dimension	(Orioua et al., 12 Sep 2025)
Probabilistic 3D regression	Negative log-likelihood of the projected Huber distribution	(Mohlin et al., 2023)
Actuarial modeling	Full conditional loss distribution $Y \mid \boldsymbol X=\boldsymbol x$ refined by a neural network	(Avanzi et al., 2024)
Financial crises	Aggregate GDP loss distribution from a Loss Distribution Approach	(Kapp et al., 2012)
Distribution grids	Technical loss estimation, reduction, or allocation in distribution networks	(Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020)

A common misconception is that “Projected Distribution Loss” names one canonical objective. In the cited work, it instead designates several non-equivalent constructions. Another recurring distinction is between a loss function used for training and a loss distribution used for forecasting or risk analysis. The former appears most clearly in image restoration, autoencoding, and 3D regression; the latter appears in actuarial, macro-financial, and power-system settings.

2. Image enhancement: Projected Distribution Loss as a perceptual objective

In image restoration, Projected Distribution Loss, or PDL, is a perceptual training objective that compares the distributions of deep features rather than their pointwise differences (Delbracio et al., 2020). The motivation is that pixel losses such as $L_1$ and $L_2$ encourage regression to the mean, and classic perceptual feature losses still compare features element-wise. Distribution-matching losses allow more flexibility, but the paper emphasizes that some are only approximate distribution matchers and adversarial objectives can produce hallucinations and are harder to optimize. PDL is positioned as an intermediate alternative: it keeps paired supervision from a reference image while relaxing strict pointwise alignment.

The formal change is to replace the direct feature norm by Wasserstein distances between projected feature distributions. With VGG feature extractor $\Phi$, feature maps $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$, projection vectors $w_j\in\mathbb{R}^m$, and projected features $\phi'_{i,j}(u)=w_j^T\phi_i(u)$, the loss is

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$

with $q=1$ and $p=1$ in all reported results (Delbracio et al., 2020). The implementation computes VGG16-conv4 features, projects them, sorts the projected samples, evaluates the 1D Wasserstein distance, sums across projections, and adds the pixel term. The paper notes that sorting is treated as fixed in backpropagation.

The paper evaluates PDL as a drop-in replacement for perceptual losses in denoising, single-image super-resolution, demosaicing, deblurring, and JPEG artifact removal. The architectures are SRN for denoising, demosaicing, and JPEG artifact removal, EDSR for super-resolution, and DsDeblur for deblurring. Training uses Adam, batch size $L_1$0, $L_1$1 iterations, VGG16-conv4 features, and DIV2K with synthetic degradations (Delbracio et al., 2020).

The reported results support the claim that the projected distribution view changes the perception–distortion trade-off. For denoising with $L_1$2, the best PDL row in Table 1 gives PSNR $L_1$3, MS-SSIM $L_1$4, LPIPS $L_1$5, NIQE $L_1$6, and FID $L_1$7. For $L_1$8 super-resolution, the best PDL result gives PSNR $L_1$9, MS-SSIM $L_2$0, LPIPS $L_2$1, NIQE $L_2$2, and FID $L_2$3. For deblurring, the best PDL result gives PSNR $L_2$4, MS-SSIM $L_2$5, LPIPS $L_2$6, NIQE $L_2$7, and FID $L_2$8 (Delbracio et al., 2020). Human preference studies reported pairwise preferences such as PDL vs no-perceptual $L_2$9, PDL vs $\Phi$0, PDL vs CTXDP $\Phi$1, and PDL vs CTXL2 $\Phi$2.

The ablation study shows that projection choice is not the dominant issue: Id, R2P, RPP, and RSP behave similarly, and the simplest channelwise marginal comparison is often as good as or better than more elaborate schemes. The paper also reports that PDL is only marginally slower than standard VGG perceptual loss, with training speeds of 0.57 steps/sec for PDL and 0.57 for $\Phi$3 perceptual loss (Delbracio et al., 2020).

3. Autoencoders and intrinsic dimension: projected reconstruction loss

In the Intrinsic Dimension Estimating Autoencoder (IDEA), the relevant notion is the projected reconstruction loss, also described as the projected loss or projected reconstruction term; the paper states that these are the same concept (Orioua et al., 12 Sep 2025). The purpose is not generic perceptual quality but the identification of the smallest effective latent dimension $\Phi$4 that still preserves reconstruction quality and should match the intrinsic dimension $\Phi$5.

IDEA uses a re-weighted double CancelOut layer. The first CancelOut layer, $\Phi$6, has trainable nonnegative weights and can suppress latent coordinates by pushing a weight to zero; the second, $\Phi$7, rescales the surviving latent features. The number of non-zero weights in $\Phi$8 is the effective latent size $\Phi$9 (Orioua et al., 12 Sep 2025). The projected term is introduced because a standard autoencoder can reconstruct well without revealing which latent dimensions are truly necessary.

The total training loss is written as a sum of four terms: reconstruction, projected reconstruction, L1 regularization, and orthogonality penalty. The projected reconstruction term is computed from $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$0, defined as “the output of the model after setting the last non-zero weight of the $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$1 layer to zero” (Orioua et al., 12 Sep 2025). Operationally, the model sees both the current reconstruction error $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$2 and the reconstruction error after removing one active dimension, $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$3. The paper characterizes elimination of a latent dimension as “a trade-off between the first two terms of the loss function.”

This projected look-ahead is the key mechanism. The paper states that it “anticipates the structure that the model will have if it decides to eliminate a dimension by setting the current $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$4 to zero” (Orioua et al., 12 Sep 2025). If the projected loss stays small, the dimension is likely redundant; if it rises sharply, the dimension is contributing real information. The orthogonality term $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$5 complements this by encouraging the latent space to become “uncorrelated.”

The empirical behavior is reported most explicitly on synthetic benchmarks. For Synthetic Legendre datasets with true dimensions $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$6, IDEA initialized with $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$7 correctly predicted $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$8. The paper reports that using $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$9 dimensions causes much larger test loss, while $w_j\in\mathbb{R}^m$0 does not improve reconstruction meaningfully; for dataset 1, $w_j\in\mathbb{R}^m$1, but $w_j\in\mathbb{R}^m$2 gives a test loss of $w_j\in\mathbb{R}^m$3, much worse than the $w_j\in\mathbb{R}^m$4 and $w_j\in\mathbb{R}^m$5 models around $w_j\in\mathbb{R}^m$6 (Orioua et al., 12 Sep 2025). The training dynamics are described in two phases: dimensions are progressively suppressed and each elimination causes a spike in the loss; then elimination stops once the trade-off is no longer favorable.

The same framework is applied to vertically resolved free-surface flow. In the simple case, IDEA predicts $w_j\in\mathbb{R}^m$7 total latent dimensions, including one dimension for water height, corresponding to a 2D manifold for the velocity profile. In the more complex case, it still predicts $w_j\in\mathbb{R}^m$8. The paper also reports that IDEA’s 2D velocity representation outperformed a truncated Legendre-moment projection that needed 4 moments in the simple case, and that in the more complex case the moment model needed at least 7 coefficients to reach comparable quality (Orioua et al., 12 Sep 2025).

4. Probabilistic 3D regression: projected Huber distribution and its loss

In monocular 3D regression, the relevant object is the negative log-likelihood of the projected Huber distribution (Mohlin et al., 2023). Here the “projected” language refers to the fact that image-plane coordinates and depth are coupled through perspective, and that a good probabilistic model should reflect this rather than predicting $w_j\in\mathbb{R}^m$9 naively.

The paper formulates seven desirable constraints for a camera-based 3D positional distribution: it should allow arbitrary variance in projected coordinates and depth independently; it should predict the projected location and depth normalized by focal length; it should assign zero probability to points behind the camera; it should assign positive probability to all physically visible points in the field of view; its negative log-likelihood should be convex in position; probability should vanish at infinity and at $\phi'_{i,j}(u)=w_j^T\phi_i(u)$0; and it should be continuous in position (Mohlin et al., 2023). The paper also proves that distributions of the form

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$1

cannot satisfy all these constraints simultaneously.

The proposed solution is the Projected Huber Distribution, composed of a projected-coordinate term conditioned on depth and a depth term with support only on $\phi'_{i,j}(u)=w_j^T\phi_i(u)$2. The projected part uses a Huber-style radial penalty, while the depth term uses

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$3

The combined distribution is $\phi'_{i,j}(u)=w_j^T\phi_i(u)$4 (Mohlin et al., 2023).

The training objective is the negative log-likelihood of this combined distribution. After a basis change to normalized coordinates,

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$5

the loss becomes

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$6

with constant terms omitted in implementation (Mohlin et al., 2023). Parameter remapping is used so that $\phi'_{i,j}(u)=w_j^T\phi_i(u)$7, $\phi'_{i,j}(u)=w_j^T\phi_i(u)$8, and $\phi'_{i,j}(u)=w_j^T\phi_i(u)$9.

This construction differs from standard MAE/MSE regression, Gaussian regression, plain Huber loss, and independent projected-and-depth probabilistic models. The paper’s emphasis is that the contribution is not just a new loss but a distributional form whose log-likelihood is tailored to camera geometry and downstream fusion (Mohlin et al., 2023). The mode used for point prediction is

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$0

Empirically, the experiments are synthetic and test focal-length variation, scale ambiguity, occlusion, and combinations of those factors. The reported conclusions are that predicted variances are well correlated with empirical squared errors; scale ambiguity raises depth uncertainty while projected uncertainty stays low; occlusion raises both projected and depth uncertainty; and the mode often outperforms MAE baselines on $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$1 and $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$2 (Mohlin et al., 2023).

5. Projected loss distributions in actuarial and macro-financial modeling

A different usage appears in actuarial science, where the object is the full conditional law of a loss variable $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$3 given covariates $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$4. The Distributional Refinement Network (DRN) defines this as the projected loss distribution, its tail behavior, and derived risk measures such as VaR (Avanzi et al., 2024). The emphasis is not on a training loss over projections, but on distributional forecasting.

The DRN begins from a standard parametric distributional regression model, typically a GLM, then refines the baseline distribution rather than replacing it. The baseline density is transformed into partitioned probability masses over intervals $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$5, with

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$6

The neural network takes the original covariates and the baseline bin masses as inputs, produces raw outputs $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$7, and refines the baseline masses multiplicatively through a softmax-type adjustment (Avanzi et al., 2024). The paper stresses that this preserves baseline structure and interpretability, because the final forecast can be decomposed into a transparent baseline component and a learned correction.

Training can be done with negative log-likelihood or the paper’s preferred joint binary cross-entropy (JBCE) objective. The regularized objective adds a KL term that keeps the DRN close to the baseline “prior,” a roughness penalty that reduces jagged density fluctuations, and a mean penalty that prevents unnecessary drift from the baseline mean when the GLM already captures the center well (Avanzi et al., 2024). On synthetic data, DRN improves test metrics from GLM values of NLL $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$8, CRPS $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$9, RMSE $q=1$0, and 90% quantile loss $q=1$1 to DRN values of NLL $q=1$2, CRPS $q=1$3, RMSE $q=1$4, and 90% quantile loss $q=1$5. On freMPL1 claim severity data, the paper reports GLM values of NLL $q=1$6, CRPS $q=1$7, RMSE $q=1$8, and 90% quantile loss $q=1$9, versus DRN values of NLL $p=1$0, CRPS $p=1$1, RMSE $p=1$2, and 90% quantile loss $p=1$3 (Avanzi et al., 2024).

The macro-financial literature uses a related but distinct framework. “Real Output Costs of Financial Crises: A Loss Distribution Approach” models crisis risk by separating frequency and severity and combining them into a global aggregate loss probability density function over five-year periods (Kapp et al., 2012). The aggregate loss is

$p=1$4

where $p=1$5 is the number of crisis events and $p=1$6 is the severity of event $p=1$7. Frequency is modeled with Poisson and Negative Binomial distributions, with the Negative Binomial preferred because the data are overdispersed; severity is fitted with Gamma, Exponential, GEV, Generalized Pareto, Log-normal, and Weibull, with Weibull used as the benchmark (Kapp et al., 2012). The reported specification uses 500,000 Monte Carlo simulations.

The resulting aggregate loss PDFs are “markedly skewed to the right.” The key headline result is that one-percent-probability global crises every five years lead to losses between $p=1$8 and $p=1$9 of world GDP, corresponding to the 99th percentile under different output-gap specifications. The 99.9th percentile reaches $L_1$00 to $L_1$01 of world GDP, while mean five-year world-GDP losses range from $L_1$02 to $L_1$03 and median losses from $L_1$04 to $L_1$05 (Kapp et al., 2012). This use of “loss distribution” is actuarial in the strict sense: it is a projected distribution of aggregate downside outcomes, not a neural-network training criterion.

6. Distribution grids: technical loss estimation, reduction, and allocation

In electrical distribution systems, the language of distribution loss refers to physical energy dissipated within the network, primarily through Joule heating, and the technical questions are how to estimate, reduce, or allocate it (Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020). This is again different from perceptual or probabilistic losses.

A first strand is voltage-only technical loss estimation. The method called voss (“Voltage to Technical Loss”) estimates the technical loss fraction of a line segment from endpoint voltage magnitudes without directly measuring currents or power flows (Bariya et al., 26 Jun 2025). For a single line segment,

$L_1$06

For a line with intervening loads, the corrected estimator is

$L_1$07

with

$L_1$08

The method is evaluated on the IEEE 13-node and 34-node feeders in OpenDSS. For line 816–822, phase A, the single-segment estimate is $L_1$09, the correction factor is $L_1$10, the multi-segment estimate is $L_1$11, and the true loss is $L_1$12, reducing the absolute error from about $L_1$13 to $L_1$14 (Bariya et al., 26 Jun 2025). The paper explicitly frames this as a practical estimator for LMIC distribution grids where conventional power-flow instrumentation is unavailable or impractical.

A second strand is coordinated loss reduction by co-optimizing Conservation Voltage Reduction (CVR) and Network Topology Reconfiguration (NTR). The objective is to minimize total active power losses over the scheduling horizon,

$L_1$15

using a mixed-integer conic program with AC branch-flow constraints, voltage-dependent ZIP loads, and radiality constraints (Fatima et al., 11 May 2026). CVR lowers nodal voltages to reduce voltage-sensitive demand and hence $L_1$16 losses, while NTR changes the feeder’s radial switching pattern to route power through lower-impedance paths and flatten branch loading. The paper argues that the interaction is multiplicative rather than additive: CVR reduces current magnitude, which enlarges the feasible region for reconfiguration, and NTR then moves those reduced currents onto lower-loss paths.

The reported numerical outcomes show that the coordinated CEDNTR case gives the lowest losses among all cases. On the IEEE 123-bus feeder, SDN has about $L_1$17 kW losses at $L_1$18 MW system load, SDNTR about $L_1$19 kW with $L_1$20 reduction, CEDN about $L_1$21 kW with $L_1$22 reduction, and CEDNTR about $L_1$23 kW with $L_1$24 reduction; the conclusion section also states a coordinated loss level of about $L_1$25 kW (Fatima et al., 11 May 2026). Maximum line loading decreases from about $L_1$26 in SDN to about $L_1$27 in CEDNTR, roughly a $L_1$28 reduction in maximum loading. The framework is tested under hourly load variation, three load compositions—constant impedance, constant current, and constant power—and with and without DERs including PV, battery energy storage, and dispatchable generation.

A third strand is loss allocation in joint transmission and distribution peer-to-peer markets. Here losses are modeled as an explicit market product: each bilateral trade $L_1$29 is accompanied by an injection $L_1$30 and a loss quantity $L_1$31, and a market operator allocates line losses through a loss-allocation matrix $L_1$32 (Moret et al., 2020). The DSO uses a linear AC approximation with distribution losses

$L_1$33

while the TSO uses

$L_1$34

The paper analyzes socialization, individual, capacity-weighted individual, and mixed allocation policies, concluding that socialization is fairer at DSO level, while individual allocation is more efficient but potentially inequitable in radial distribution grids (Moret et al., 2020).

Taken together, these power-system papers show that “distribution loss” can denote a measurable physical quantity, an optimization target, or an allocable market object. This suggests that in infrastructure research the phrase is anchored in network physics and operational planning rather than in representation learning or probabilistic perception.PY"ERR\", e)\nPY"]' to=shell 彩神争霸如何 code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?search_query=all:%22Projected+Distribution+Loss%22&start=0&max_results=5'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(data[:1200])\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?search_query=all:%22Projected+Distribution+Loss%22&start=0&max_results=5'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(data[:1200])\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nids=['(Delbracio et al., 2020)','(Orioua et al., 12 Sep 2025)','(Mohlin et al., 2023)','(Avanzi et al., 2024)','(Fatima et al., 11 May 2026)']\nfor aid in ids:\n url=f'http://export.arxiv.org/api/query?id_list={aid}'\n try:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n title=''\n if '<title>' in data:\n parts=data.split('<entry>')\n if len(parts)>1:\n seg=parts[1]\n t=seg.split('<title>')[1].split('</title>')[0].strip().replace('\n',' ')\n title=t\n print(aid, title)\n except Exception as e:\n print('ERR', aid, e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?id_list=([2012.09289](/papers/2012.09289))'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(repr(data[:500]))\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","echo hi"]} “Projected Distribution Loss” is not a single standardized construct across arXiv literature. The expression and closely related variants are used for several technically distinct objects: a perceptual training objective for image enhancement based on projected feature distributions; a projected reconstruction term for intrinsic-dimension estimation in autoencoders; a camera-aware negative log-likelihood for probabilistic 3D regression; a forecast of the full conditional law of losses in actuarial modeling; a compound aggregate loss distribution for financial crises; and the estimation, reduction, or allocation of technical losses in electrical distribution networks (Delbracio et al., 2020, Orioua et al., 12 Sep 2025, Mohlin et al., 2023, Avanzi et al., 2024, Kapp et al., 2012, Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020). This suggests that the term is field-dependent: in some papers it denotes a training loss built from projected representations, while in others it denotes a projected or estimated distribution of losses, or physical loss in distribution systems.

1. Terminological scope and recurrent idea

Across the cited literature, the shared motif is the use of a projection, refinement, or conditional distribution to represent uncertainty or loss more faithfully than a simple point estimate. The technical object, however, changes by domain.

Domain	Meaning in the source literature	Source
Image enhancement	Perceptual loss using aggregated 1D-Wasserstein distances between projected CNN features	(Delbracio et al., 2020)
Intrinsic-dimension estimation	Projected reconstruction loss after removing one active latent dimension	(Orioua et al., 12 Sep 2025)
Probabilistic 3D regression	Negative log-likelihood of the projected Huber distribution	(Mohlin et al., 2023)
Actuarial modeling	Full conditional loss distribution $Y \mid \boldsymbol X=\boldsymbol x$ refined by a neural network	(Avanzi et al., 2024)
Financial crises	Aggregate GDP loss distribution from a Loss Distribution Approach	(Kapp et al., 2012)
Distribution grids	Technical loss estimation, reduction, or allocation in distribution networks	(Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020)

A common misconception is that “Projected Distribution Loss” names one canonical objective. In the cited work, it instead designates several non-equivalent constructions. Another recurring distinction is between a loss function used for training and a loss distribution used for forecasting or risk analysis. The former appears most clearly in image restoration, autoencoding, and 3D regression; the latter appears in actuarial, macro-financial, and power-system settings.

2. Image enhancement: Projected Distribution Loss as a perceptual objective

In image restoration, Projected Distribution Loss, or PDL, is a perceptual training objective that compares the distributions of deep features rather than their pointwise differences (Delbracio et al., 2020). The motivation is that pixel losses such as $L_1$ and $L_2$ encourage regression to the mean, and classic perceptual feature losses still compare features element-wise. Distribution-matching losses allow more flexibility, but the paper emphasizes that some are only approximate distribution matchers and adversarial objectives can produce hallucinations and are harder to optimize. PDL is positioned as an intermediate alternative: it keeps paired supervision from a reference image while relaxing strict pointwise alignment.

The formal change is to replace the direct feature norm by Wasserstein distances between projected feature distributions. With VGG feature extractor $\Phi$, feature maps $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$, projection vectors $w_j\in\mathbb{R}^m$, and projected features $\phi'_{i,j}(u)=w_j^T\phi_i(u)$, the loss is

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$

with $q=1$ and $p=1$ in all reported results (Delbracio et al., 2020). The implementation computes VGG16-conv4 features, projects them, sorts the projected samples, evaluates the 1D Wasserstein distance, sums across projections, and adds the pixel term. The paper notes that sorting is treated as fixed in backpropagation.

The paper evaluates PDL as a drop-in replacement for perceptual losses in denoising, single-image super-resolution, demosaicing, deblurring, and JPEG artifact removal. The architectures are SRN for denoising, demosaicing, and JPEG artifact removal, EDSR for super-resolution, and DsDeblur for deblurring. Training uses Adam, batch size $L_1$0, $L_1$1 iterations, VGG16-conv4 features, and DIV2K with synthetic degradations (Delbracio et al., 2020).

The reported results support the claim that the projected distribution view changes the perception–distortion trade-off. For denoising with $L_1$2, the best PDL row in Table 1 gives PSNR $L_1$3, MS-SSIM $L_1$4, LPIPS $L_1$5, NIQE $L_1$6, and FID $L_1$7. For $L_1$8 super-resolution, the best PDL result gives PSNR $L_1$9, MS-SSIM $L_2$0, LPIPS $L_2$1, NIQE $L_2$2, and FID $L_2$3. For deblurring, the best PDL result gives PSNR $L_2$4, MS-SSIM $L_2$5, LPIPS $L_2$6, NIQE $L_2$7, and FID $L_2$8 (Delbracio et al., 2020). Human preference studies reported pairwise preferences such as PDL vs no-perceptual $L_2$9, PDL vs $\Phi$0, PDL vs CTXDP $\Phi$1, and PDL vs CTXL2 $\Phi$2.

The ablation study shows that projection choice is not the dominant issue: Id, R2P, RPP, and RSP behave similarly, and the simplest channelwise marginal comparison is often as good as or better than more elaborate schemes. The paper also reports that PDL is only marginally slower than standard VGG perceptual loss, with training speeds of 0.57 steps/sec for PDL and 0.57 for $\Phi$3 perceptual loss (Delbracio et al., 2020).

3. Autoencoders and intrinsic dimension: projected reconstruction loss

In the Intrinsic Dimension Estimating Autoencoder (IDEA), the relevant notion is the projected reconstruction loss, also described as the projected loss or projected reconstruction term; the paper states that these are the same concept (Orioua et al., 12 Sep 2025). The purpose is not generic perceptual quality but the identification of the smallest effective latent dimension $\Phi$4 that still preserves reconstruction quality and should match the intrinsic dimension $\Phi$5.

IDEA uses a re-weighted double CancelOut layer. The first CancelOut layer, $\Phi$6, has trainable nonnegative weights and can suppress latent coordinates by pushing a weight to zero; the second, $\Phi$7, rescales the surviving latent features. The number of non-zero weights in $\Phi$8 is the effective latent size $\Phi$9 (Orioua et al., 12 Sep 2025). The projected term is introduced because a standard autoencoder can reconstruct well without revealing which latent dimensions are truly necessary.

The total training loss is written as a sum of four terms: reconstruction, projected reconstruction, L1 regularization, and orthogonality penalty. The projected reconstruction term is computed from $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$0, defined as “the output of the model after setting the last non-zero weight of the $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$1 layer to zero” (Orioua et al., 12 Sep 2025). Operationally, the model sees both the current reconstruction error $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$2 and the reconstruction error after removing one active dimension, $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$3. The paper characterizes elimination of a latent dimension as “a trade-off between the first two terms of the loss function.”

This projected look-ahead is the key mechanism. The paper states that it “anticipates the structure that the model will have if it decides to eliminate a dimension by setting the current $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$4 to zero” (Orioua et al., 12 Sep 2025). If the projected loss stays small, the dimension is likely redundant; if it rises sharply, the dimension is contributing real information. The orthogonality term $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$5 complements this by encouraging the latent space to become “uncorrelated.”

The empirical behavior is reported most explicitly on synthetic benchmarks. For Synthetic Legendre datasets with true dimensions $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$6, IDEA initialized with $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$7 correctly predicted $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$8. The paper reports that using $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$9 dimensions causes much larger test loss, while $w_j\in\mathbb{R}^m$0 does not improve reconstruction meaningfully; for dataset 1, $w_j\in\mathbb{R}^m$1, but $w_j\in\mathbb{R}^m$2 gives a test loss of $w_j\in\mathbb{R}^m$3, much worse than the $w_j\in\mathbb{R}^m$4 and $w_j\in\mathbb{R}^m$5 models around $w_j\in\mathbb{R}^m$6 (Orioua et al., 12 Sep 2025). The training dynamics are described in two phases: dimensions are progressively suppressed and each elimination causes a spike in the loss; then elimination stops once the trade-off is no longer favorable.

The same framework is applied to vertically resolved free-surface flow. In the simple case, IDEA predicts $w_j\in\mathbb{R}^m$7 total latent dimensions, including one dimension for water height, corresponding to a 2D manifold for the velocity profile. In the more complex case, it still predicts $w_j\in\mathbb{R}^m$8. The paper also reports that IDEA’s 2D velocity representation outperformed a truncated Legendre-moment projection that needed 4 moments in the simple case, and that in the more complex case the moment model needed at least 7 coefficients to reach comparable quality (Orioua et al., 12 Sep 2025).

4. Probabilistic 3D regression: projected Huber distribution and its loss

In monocular 3D regression, the relevant object is the negative log-likelihood of the projected Huber distribution (Mohlin et al., 2023). Here the “projected” language refers to the fact that image-plane coordinates and depth are coupled through perspective, and that a good probabilistic model should reflect this rather than predicting $w_j\in\mathbb{R}^m$9 naively.

The paper formulates seven desirable constraints for a camera-based 3D positional distribution: it should allow arbitrary variance in projected coordinates and depth independently; it should predict the projected location and depth normalized by focal length; it should assign zero probability to points behind the camera; it should assign positive probability to all physically visible points in the field of view; its negative log-likelihood should be convex in position; probability should vanish at infinity and at $\phi'_{i,j}(u)=w_j^T\phi_i(u)$0; and it should be continuous in position (Mohlin et al., 2023). The paper also proves that distributions of the form

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$1

cannot satisfy all these constraints simultaneously.

The proposed solution is the Projected Huber Distribution, composed of a projected-coordinate term conditioned on depth and a depth term with support only on $\phi'_{i,j}(u)=w_j^T\phi_i(u)$2. The projected part uses a Huber-style radial penalty, while the depth term uses

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$3

The combined distribution is $\phi'_{i,j}(u)=w_j^T\phi_i(u)$4 (Mohlin et al., 2023).

The training objective is the negative log-likelihood of this combined distribution. After a basis change to normalized coordinates,

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$5

the loss becomes

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$6

with constant terms omitted in implementation (Mohlin et al., 2023). Parameter remapping is used so that $\phi'_{i,j}(u)=w_j^T\phi_i(u)$7, $\phi'_{i,j}(u)=w_j^T\phi_i(u)$8, and $\phi'_{i,j}(u)=w_j^T\phi_i(u)$9.

This construction differs from standard MAE/MSE regression, Gaussian regression, plain Huber loss, and independent projected-and-depth probabilistic models. The paper’s emphasis is that the contribution is not just a new loss but a distributional form whose log-likelihood is tailored to camera geometry and downstream fusion (Mohlin et al., 2023). The mode used for point prediction is

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$0

Empirically, the experiments are synthetic and test focal-length variation, scale ambiguity, occlusion, and combinations of those factors. The reported conclusions are that predicted variances are well correlated with empirical squared errors; scale ambiguity raises depth uncertainty while projected uncertainty stays low; occlusion raises both projected and depth uncertainty; and the mode often outperforms MAE baselines on $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$1 and $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$2 (Mohlin et al., 2023).

5. Projected loss distributions in actuarial and macro-financial modeling

A different usage appears in actuarial science, where the object is the full conditional law of a loss variable $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$3 given covariates $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$4. The Distributional Refinement Network (DRN) defines this as the projected loss distribution, its tail behavior, and derived risk measures such as VaR (Avanzi et al., 2024). The emphasis is not on a training loss over projections, but on distributional forecasting.

The DRN begins from a standard parametric distributional regression model, typically a GLM, then refines the baseline distribution rather than replacing it. The baseline density is transformed into partitioned probability masses over intervals $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$5, with

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$6

The neural network takes the original covariates and the baseline bin masses as inputs, produces raw outputs $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$7, and refines the baseline masses multiplicatively through a softmax-type adjustment (Avanzi et al., 2024). The paper stresses that this preserves baseline structure and interpretability, because the final forecast can be decomposed into a transparent baseline component and a learned correction.

Training can be done with negative log-likelihood or the paper’s preferred joint binary cross-entropy (JBCE) objective. The regularized objective adds a KL term that keeps the DRN close to the baseline “prior,” a roughness penalty that reduces jagged density fluctuations, and a mean penalty that prevents unnecessary drift from the baseline mean when the GLM already captures the center well (Avanzi et al., 2024). On synthetic data, DRN improves test metrics from GLM values of NLL $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$8, CRPS $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$9, RMSE $q=1$0, and 90% quantile loss $q=1$1 to DRN values of NLL $q=1$2, CRPS $q=1$3, RMSE $q=1$4, and 90% quantile loss $q=1$5. On freMPL1 claim severity data, the paper reports GLM values of NLL $q=1$6, CRPS $q=1$7, RMSE $q=1$8, and 90% quantile loss $q=1$9, versus DRN values of NLL $p=1$0, CRPS $p=1$1, RMSE $p=1$2, and 90% quantile loss $p=1$3 (Avanzi et al., 2024).

The macro-financial literature uses a related but distinct framework. “Real Output Costs of Financial Crises: A Loss Distribution Approach” models crisis risk by separating frequency and severity and combining them into a global aggregate loss probability density function over five-year periods (Kapp et al., 2012). The aggregate loss is

$p=1$4

where $p=1$5 is the number of crisis events and $p=1$6 is the severity of event $p=1$7. Frequency is modeled with Poisson and Negative Binomial distributions, with the Negative Binomial preferred because the data are overdispersed; severity is fitted with Gamma, Exponential, GEV, Generalized Pareto, Log-normal, and Weibull, with Weibull used as the benchmark (Kapp et al., 2012). The reported specification uses 500,000 Monte Carlo simulations.

The resulting aggregate loss PDFs are “markedly skewed to the right.” The key headline result is that one-percent-probability global crises every five years lead to losses between $p=1$8 and $p=1$9 of world GDP, corresponding to the 99th percentile under different output-gap specifications. The 99.9th percentile reaches $L_1$00 to $L_1$01 of world GDP, while mean five-year world-GDP losses range from $L_1$02 to $L_1$03 and median losses from $L_1$04 to $L_1$05 (Kapp et al., 2012). This use of “loss distribution” is actuarial in the strict sense: it is a projected distribution of aggregate downside outcomes, not a neural-network training criterion.

6. Distribution grids: technical loss estimation, reduction, and allocation

In electrical distribution systems, the language of distribution loss refers to physical energy dissipated within the network, primarily through Joule heating, and the technical questions are how to estimate, reduce, or allocate it (Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020). This is again different from perceptual or probabilistic losses.

A first strand is voltage-only technical loss estimation. The method called voss (“Voltage to Technical Loss”) estimates the technical loss fraction of a line segment from endpoint voltage magnitudes without directly measuring currents or power flows (Bariya et al., 26 Jun 2025). For a single line segment,

$L_1$06

For a line with intervening loads, the corrected estimator is

$L_1$07

with

$L_1$08

The method is evaluated on the IEEE 13-node and 34-node feeders in OpenDSS. For line 816–822, phase A, the single-segment estimate is $L_1$09, the correction factor is $L_1$10, the multi-segment estimate is $L_1$11, and the true loss is $L_1$12, reducing the absolute error from about $L_1$13 to $L_1$14 (Bariya et al., 26 Jun 2025). The paper explicitly frames this as a practical estimator for LMIC distribution grids where conventional power-flow instrumentation is unavailable or impractical.

A second strand is coordinated loss reduction by co-optimizing Conservation Voltage Reduction (CVR) and Network Topology Reconfiguration (NTR). The objective is to minimize total active power losses over the scheduling horizon,

$L_1$15

using a mixed-integer conic program with AC branch-flow constraints, voltage-dependent ZIP loads, and radiality constraints (Fatima et al., 11 May 2026). CVR lowers nodal voltages to reduce voltage-sensitive demand and hence $L_1$16 losses, while NTR changes the feeder’s radial switching pattern to route power through lower-impedance paths and flatten branch loading. The paper argues that the interaction is multiplicative rather than additive: CVR reduces current magnitude, which enlarges the feasible region for reconfiguration, and NTR then moves those reduced currents onto lower-loss paths.

The reported numerical outcomes show that the coordinated CEDNTR case gives the lowest losses among all cases. On the IEEE 123-bus feeder, SDN has about $L_1$17 kW losses at $L_1$18 MW system load, SDNTR about $L_1$19 kW with $L_1$20 reduction, CEDN about $L_1$21 kW with $L_1$22 reduction, and CEDNTR about $L_1$23 kW with $L_1$24 reduction; the conclusion section also states a coordinated loss level of about $L_1$25 kW (Fatima et al., 11 May 2026). Maximum line loading decreases from about $L_1$26 in SDN to about $L_1$27 in CEDNTR, roughly a $L_1$28 reduction in maximum loading. The framework is tested under hourly load variation, three load compositions—constant impedance, constant current, and constant power—and with and without DERs including PV, battery energy storage, and dispatchable generation.

A third strand is loss allocation in joint transmission and distribution peer-to-peer markets. Here losses are modeled as an explicit market product: each bilateral trade $L_1$29 is accompanied by an injection $L_1$30 and a loss quantity $L_1$31, and a market operator allocates line losses through a loss-allocation matrix $L_1$32 (Moret et al., 2020). The DSO uses a linear AC approximation with distribution losses

$L_1$33

while the TSO uses

$L_1$34

The paper analyzes socialization, individual, capacity-weighted individual, and mixed allocation policies, concluding that socialization is fairer at DSO level, while individual allocation is more efficient but potentially inequitable in radial distribution grids (Moret et al., 2020).

Taken together, these power-system papers show that “distribution loss” can denote a measurable physical quantity, an optimization target, or an allocable market object. This suggests that in infrastructure research the phrase is anchored in network physics and operational planning rather than in representation learning or probabilistic perception.\nimport urllib.request\nurl="ERR\", e)\nPY"]' to=shell 彩神争霸如何 code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?search_query=all:%22Projected+Distribution+Loss%22&start=0&max_results=5'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(data[:1200])\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?search_query=all:%22Projected+Distribution+Loss%22&start=0&max_results=5'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(data[:1200])\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nids=['(Delbracio et al., 2020)','(Orioua et al., 12 Sep 2025)','(Mohlin et al., 2023)','(Avanzi et al., 2024)','(Fatima et al., 11 May 2026)']\nfor aid in ids:\n url=f'http://export.arxiv.org/api/query?id_list={aid}'\n try:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n title=''\n if '<title>' in data:\n parts=data.split('<entry>')\n if len(parts)>1:\n seg=parts[1]\n t=seg.split('<title>')[1].split('</title>')[0].strip().replace('\n',' ')\n title=t\n print(aid, title)\n except Exception as e:\n print('ERR', aid, e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?id_list=([2012.09289](/papers/2012.09289))'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(repr(data[:500]))\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","echo hi"]} “Projected Distribution Loss” is not a single standardized construct across arXiv literature. The expression and closely related variants are used for several technically distinct objects: a perceptual training objective for image enhancement based on projected feature distributions; a projected reconstruction term for intrinsic-dimension estimation in autoencoders; a camera-aware negative log-likelihood for probabilistic 3D regression; a forecast of the full conditional law of losses in actuarial modeling; a compound aggregate loss distribution for financial crises; and the estimation, reduction, or allocation of technical losses in electrical distribution networks (Delbracio et al., 2020, Orioua et al., 12 Sep 2025, Mohlin et al., 2023, Avanzi et al., 2024, Kapp et al., 2012, Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020). This suggests that the term is field-dependent: in some papers it denotes a training loss built from projected representations, while in others it denotes a projected or estimated distribution of losses, or physical loss in distribution systems.

1. Terminological scope and recurrent idea

Across the cited literature, the shared motif is the use of a projection, refinement, or conditional distribution to represent uncertainty or loss more faithfully than a simple point estimate. The technical object, however, changes by domain.

Domain	Meaning in the source literature	Source
Image enhancement	Perceptual loss using aggregated 1D-Wasserstein distances between projected CNN features	(Delbracio et al., 2020)
Intrinsic-dimension estimation	Projected reconstruction loss after removing one active latent dimension	(Orioua et al., 12 Sep 2025)
Probabilistic 3D regression	Negative log-likelihood of the projected Huber distribution	(Mohlin et al., 2023)
Actuarial modeling	Full conditional loss distribution $Y \mid \boldsymbol X=\boldsymbol x$ refined by a neural network	(Avanzi et al., 2024)
Financial crises	Aggregate GDP loss distribution from a Loss Distribution Approach	(Kapp et al., 2012)
Distribution grids	Technical loss estimation, reduction, or allocation in distribution networks	(Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020)

A common misconception is that “Projected Distribution Loss” names one canonical objective. In the cited work, it instead designates several non-equivalent constructions. Another recurring distinction is between a loss function used for training and a loss distribution used for forecasting or risk analysis. The former appears most clearly in image restoration, autoencoding, and 3D regression; the latter appears in actuarial, macro-financial, and power-system settings.

2. Image enhancement: Projected Distribution Loss as a perceptual objective

In image restoration, Projected Distribution Loss, or PDL, is a perceptual training objective that compares the distributions of deep features rather than their pointwise differences (Delbracio et al., 2020). The motivation is that pixel losses such as $L_1$ and $L_2$ encourage regression to the mean, and classic perceptual feature losses still compare features element-wise. Distribution-matching losses allow more flexibility, but the paper emphasizes that some are only approximate distribution matchers and adversarial objectives can produce hallucinations and are harder to optimize. PDL is positioned as an intermediate alternative: it keeps paired supervision from a reference image while relaxing strict pointwise alignment.

The formal change is to replace the direct feature norm by Wasserstein distances between projected feature distributions. With VGG feature extractor $\Phi$, feature maps $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$, projection vectors $w_j\in\mathbb{R}^m$, and projected features $\phi'_{i,j}(u)=w_j^T\phi_i(u)$, the loss is

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$

with $q=1$ and $p=1$ in all reported results (Delbracio et al., 2020). The implementation computes VGG16-conv4 features, projects them, sorts the projected samples, evaluates the 1D Wasserstein distance, sums across projections, and adds the pixel term. The paper notes that sorting is treated as fixed in backpropagation.

The paper evaluates PDL as a drop-in replacement for perceptual losses in denoising, single-image super-resolution, demosaicing, deblurring, and JPEG artifact removal. The architectures are SRN for denoising, demosaicing, and JPEG artifact removal, EDSR for super-resolution, and DsDeblur for deblurring. Training uses Adam, batch size $L_1$0, $L_1$1 iterations, VGG16-conv4 features, and DIV2K with synthetic degradations (Delbracio et al., 2020).

The reported results support the claim that the projected distribution view changes the perception–distortion trade-off. For denoising with $L_1$2, the best PDL row in Table 1 gives PSNR $L_1$3, MS-SSIM $L_1$4, LPIPS $L_1$5, NIQE $L_1$6, and FID $L_1$7. For $L_1$8 super-resolution, the best PDL result gives PSNR $L_1$9, MS-SSIM $L_2$0, LPIPS $L_2$1, NIQE $L_2$2, and FID $L_2$3. For deblurring, the best PDL result gives PSNR $L_2$4, MS-SSIM $L_2$5, LPIPS $L_2$6, NIQE $L_2$7, and FID $L_2$8 (Delbracio et al., 2020). Human preference studies reported pairwise preferences such as PDL vs no-perceptual $L_2$9, PDL vs $\Phi$0, PDL vs CTXDP $\Phi$1, and PDL vs CTXL2 $\Phi$2.

The ablation study shows that projection choice is not the dominant issue: Id, R2P, RPP, and RSP behave similarly, and the simplest channelwise marginal comparison is often as good as or better than more elaborate schemes. The paper also reports that PDL is only marginally slower than standard VGG perceptual loss, with training speeds of 0.57 steps/sec for PDL and 0.57 for $\Phi$3 perceptual loss (Delbracio et al., 2020).

3. Autoencoders and intrinsic dimension: projected reconstruction loss

In the Intrinsic Dimension Estimating Autoencoder (IDEA), the relevant notion is the projected reconstruction loss, also described as the projected loss or projected reconstruction term; the paper states that these are the same concept (Orioua et al., 12 Sep 2025). The purpose is not generic perceptual quality but the identification of the smallest effective latent dimension $\Phi$4 that still preserves reconstruction quality and should match the intrinsic dimension $\Phi$5.

IDEA uses a re-weighted double CancelOut layer. The first CancelOut layer, $\Phi$6, has trainable nonnegative weights and can suppress latent coordinates by pushing a weight to zero; the second, $\Phi$7, rescales the surviving latent features. The number of non-zero weights in $\Phi$8 is the effective latent size $\Phi$9 (Orioua et al., 12 Sep 2025). The projected term is introduced because a standard autoencoder can reconstruct well without revealing which latent dimensions are truly necessary.

The total training loss is written as a sum of four terms: reconstruction, projected reconstruction, L1 regularization, and orthogonality penalty. The projected reconstruction term is computed from $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$0, defined as “the output of the model after setting the last non-zero weight of the $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$1 layer to zero” (Orioua et al., 12 Sep 2025). Operationally, the model sees both the current reconstruction error $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$2 and the reconstruction error after removing one active dimension, $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$3. The paper characterizes elimination of a latent dimension as “a trade-off between the first two terms of the loss function.”

This projected look-ahead is the key mechanism. The paper states that it “anticipates the structure that the model will have if it decides to eliminate a dimension by setting the current $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$4 to zero” (Orioua et al., 12 Sep 2025). If the projected loss stays small, the dimension is likely redundant; if it rises sharply, the dimension is contributing real information. The orthogonality term $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$5 complements this by encouraging the latent space to become “uncorrelated.”

The empirical behavior is reported most explicitly on synthetic benchmarks. For Synthetic Legendre datasets with true dimensions $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$6, IDEA initialized with $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$7 correctly predicted $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$8. The paper reports that using $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$9 dimensions causes much larger test loss, while $w_j\in\mathbb{R}^m$0 does not improve reconstruction meaningfully; for dataset 1, $w_j\in\mathbb{R}^m$1, but $w_j\in\mathbb{R}^m$2 gives a test loss of $w_j\in\mathbb{R}^m$3, much worse than the $w_j\in\mathbb{R}^m$4 and $w_j\in\mathbb{R}^m$5 models around $w_j\in\mathbb{R}^m$6 (Orioua et al., 12 Sep 2025). The training dynamics are described in two phases: dimensions are progressively suppressed and each elimination causes a spike in the loss; then elimination stops once the trade-off is no longer favorable.

The same framework is applied to vertically resolved free-surface flow. In the simple case, IDEA predicts $w_j\in\mathbb{R}^m$7 total latent dimensions, including one dimension for water height, corresponding to a 2D manifold for the velocity profile. In the more complex case, it still predicts $w_j\in\mathbb{R}^m$8. The paper also reports that IDEA’s 2D velocity representation outperformed a truncated Legendre-moment projection that needed 4 moments in the simple case, and that in the more complex case the moment model needed at least 7 coefficients to reach comparable quality (Orioua et al., 12 Sep 2025).

4. Probabilistic 3D regression: projected Huber distribution and its loss

In monocular 3D regression, the relevant object is the negative log-likelihood of the projected Huber distribution (Mohlin et al., 2023). Here the “projected” language refers to the fact that image-plane coordinates and depth are coupled through perspective, and that a good probabilistic model should reflect this rather than predicting $w_j\in\mathbb{R}^m$9 naively.

The paper formulates seven desirable constraints for a camera-based 3D positional distribution: it should allow arbitrary variance in projected coordinates and depth independently; it should predict the projected location and depth normalized by focal length; it should assign zero probability to points behind the camera; it should assign positive probability to all physically visible points in the field of view; its negative log-likelihood should be convex in position; probability should vanish at infinity and at $\phi'_{i,j}(u)=w_j^T\phi_i(u)$0; and it should be continuous in position (Mohlin et al., 2023). The paper also proves that distributions of the form

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$1

cannot satisfy all these constraints simultaneously.

The proposed solution is the Projected Huber Distribution, composed of a projected-coordinate term conditioned on depth and a depth term with support only on $\phi'_{i,j}(u)=w_j^T\phi_i(u)$2. The projected part uses a Huber-style radial penalty, while the depth term uses

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$3

The combined distribution is $\phi'_{i,j}(u)=w_j^T\phi_i(u)$4 (Mohlin et al., 2023).

The training objective is the negative log-likelihood of this combined distribution. After a basis change to normalized coordinates,

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$5

the loss becomes

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$6

with constant terms omitted in implementation (Mohlin et al., 2023). Parameter remapping is used so that $\phi'_{i,j}(u)=w_j^T\phi_i(u)$7, $\phi'_{i,j}(u)=w_j^T\phi_i(u)$8, and $\phi'_{i,j}(u)=w_j^T\phi_i(u)$9.

This construction differs from standard MAE/MSE regression, Gaussian regression, plain Huber loss, and independent projected-and-depth probabilistic models. The paper’s emphasis is that the contribution is not just a new loss but a distributional form whose log-likelihood is tailored to camera geometry and downstream fusion (Mohlin et al., 2023). The mode used for point prediction is

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$0

Empirically, the experiments are synthetic and test focal-length variation, scale ambiguity, occlusion, and combinations of those factors. The reported conclusions are that predicted variances are well correlated with empirical squared errors; scale ambiguity raises depth uncertainty while projected uncertainty stays low; occlusion raises both projected and depth uncertainty; and the mode often outperforms MAE baselines on $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$1 and $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$2 (Mohlin et al., 2023).

5. Projected loss distributions in actuarial and macro-financial modeling

A different usage appears in actuarial science, where the object is the full conditional law of a loss variable $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$3 given covariates $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$4. The Distributional Refinement Network (DRN) defines this as the projected loss distribution, its tail behavior, and derived risk measures such as VaR (Avanzi et al., 2024). The emphasis is not on a training loss over projections, but on distributional forecasting.

The DRN begins from a standard parametric distributional regression model, typically a GLM, then refines the baseline distribution rather than replacing it. The baseline density is transformed into partitioned probability masses over intervals $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$5, with

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$6

The neural network takes the original covariates and the baseline bin masses as inputs, produces raw outputs $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$7, and refines the baseline masses multiplicatively through a softmax-type adjustment (Avanzi et al., 2024). The paper stresses that this preserves baseline structure and interpretability, because the final forecast can be decomposed into a transparent baseline component and a learned correction.

Training can be done with negative log-likelihood or the paper’s preferred joint binary cross-entropy (JBCE) objective. The regularized objective adds a KL term that keeps the DRN close to the baseline “prior,” a roughness penalty that reduces jagged density fluctuations, and a mean penalty that prevents unnecessary drift from the baseline mean when the GLM already captures the center well (Avanzi et al., 2024). On synthetic data, DRN improves test metrics from GLM values of NLL $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$8, CRPS $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$9, RMSE $q=1$0, and 90% quantile loss $q=1$1 to DRN values of NLL $q=1$2, CRPS $q=1$3, RMSE $q=1$4, and 90% quantile loss $q=1$5. On freMPL1 claim severity data, the paper reports GLM values of NLL $q=1$6, CRPS $q=1$7, RMSE $q=1$8, and 90% quantile loss $q=1$9, versus DRN values of NLL $p=1$0, CRPS $p=1$1, RMSE $p=1$2, and 90% quantile loss $p=1$3 (Avanzi et al., 2024).

The macro-financial literature uses a related but distinct framework. “Real Output Costs of Financial Crises: A Loss Distribution Approach” models crisis risk by separating frequency and severity and combining them into a global aggregate loss probability density function over five-year periods (Kapp et al., 2012). The aggregate loss is

$p=1$4

where $p=1$5 is the number of crisis events and $p=1$6 is the severity of event $p=1$7. Frequency is modeled with Poisson and Negative Binomial distributions, with the Negative Binomial preferred because the data are overdispersed; severity is fitted with Gamma, Exponential, GEV, Generalized Pareto, Log-normal, and Weibull, with Weibull used as the benchmark (Kapp et al., 2012). The reported specification uses 500,000 Monte Carlo simulations.

The resulting aggregate loss PDFs are “markedly skewed to the right.” The key headline result is that one-percent-probability global crises every five years lead to losses between $p=1$8 and $p=1$9 of world GDP, corresponding to the 99th percentile under different output-gap specifications. The 99.9th percentile reaches $L_1$00 to $L_1$01 of world GDP, while mean five-year world-GDP losses range from $L_1$02 to $L_1$03 and median losses from $L_1$04 to $L_1$05 (Kapp et al., 2012). This use of “loss distribution” is actuarial in the strict sense: it is a projected distribution of aggregate downside outcomes, not a neural-network training criterion.

6. Distribution grids: technical loss estimation, reduction, and allocation

In electrical distribution systems, the language of distribution loss refers to physical energy dissipated within the network, primarily through Joule heating, and the technical questions are how to estimate, reduce, or allocate it (Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020). This is again different from perceptual or probabilistic losses.

A first strand is voltage-only technical loss estimation. The method called voss (“Voltage to Technical Loss”) estimates the technical loss fraction of a line segment from endpoint voltage magnitudes without directly measuring currents or power flows (Bariya et al., 26 Jun 2025). For a single line segment,

$L_1$06

For a line with intervening loads, the corrected estimator is

$L_1$07

with

$L_1$08

The method is evaluated on the IEEE 13-node and 34-node feeders in OpenDSS. For line 816–822, phase A, the single-segment estimate is $L_1$09, the correction factor is $L_1$10, the multi-segment estimate is $L_1$11, and the true loss is $L_1$12, reducing the absolute error from about $L_1$13 to $L_1$14 (Bariya et al., 26 Jun 2025). The paper explicitly frames this as a practical estimator for LMIC distribution grids where conventional power-flow instrumentation is unavailable or impractical.

A second strand is coordinated loss reduction by co-optimizing Conservation Voltage Reduction (CVR) and Network Topology Reconfiguration (NTR). The objective is to minimize total active power losses over the scheduling horizon,

$L_1$15

using a mixed-integer conic program with AC branch-flow constraints, voltage-dependent ZIP loads, and radiality constraints (Fatima et al., 11 May 2026). CVR lowers nodal voltages to reduce voltage-sensitive demand and hence $L_1$16 losses, while NTR changes the feeder’s radial switching pattern to route power through lower-impedance paths and flatten branch loading. The paper argues that the interaction is multiplicative rather than additive: CVR reduces current magnitude, which enlarges the feasible region for reconfiguration, and NTR then moves those reduced currents onto lower-loss paths.

The reported numerical outcomes show that the coordinated CEDNTR case gives the lowest losses among all cases. On the IEEE 123-bus feeder, SDN has about $L_1$17 kW losses at $L_1$18 MW system load, SDNTR about $L_1$19 kW with $L_1$20 reduction, CEDN about $L_1$21 kW with $L_1$22 reduction, and CEDNTR about $L_1$23 kW with $L_1$24 reduction; the conclusion section also states a coordinated loss level of about $L_1$25 kW (Fatima et al., 11 May 2026). Maximum line loading decreases from about $L_1$26 in SDN to about $L_1$27 in CEDNTR, roughly a $L_1$28 reduction in maximum loading. The framework is tested under hourly load variation, three load compositions—constant impedance, constant current, and constant power—and with and without DERs including PV, battery energy storage, and dispatchable generation.

A third strand is loss allocation in joint transmission and distribution peer-to-peer markets. Here losses are modeled as an explicit market product: each bilateral trade $L_1$29 is accompanied by an injection $L_1$30 and a loss quantity $L_1$31, and a market operator allocates line losses through a loss-allocation matrix $L_1$32 (Moret et al., 2020). The DSO uses a linear AC approximation with distribution losses

$L_1$33

while the TSO uses

$L_1$34

The paper analyzes socialization, individual, capacity-weighted individual, and mixed allocation policies, concluding that socialization is fairer at DSO level, while individual allocation is more efficient but potentially inequitable in radial distribution grids (Moret et al., 2020).

Taken together, these power-system papers show that “distribution loss” can denote a measurable physical quantity, an optimization target, or an allocable market object. This suggests that in infrastructure research the phrase is anchored in network physics and operational planning rather than in representation learning or probabilistic perception.http://export.arxiv.org/api/query?search_query=all:%22Projected+Distribution+Loss%22&start=0&max_results=5"ERR\", e)\nPY"]' to=shell 彩神争霸如何 code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?search_query=all:%22Projected+Distribution+Loss%22&start=0&max_results=5'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(data[:1200])\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?search_query=all:%22Projected+Distribution+Loss%22&start=0&max_results=5'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(data[:1200])\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nids=['(Delbracio et al., 2020)','(Orioua et al., 12 Sep 2025)','(Mohlin et al., 2023)','(Avanzi et al., 2024)','(Fatima et al., 11 May 2026)']\nfor aid in ids:\n url=f'http://export.arxiv.org/api/query?id_list={aid}'\n try:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n title=''\n if '<title>' in data:\n parts=data.split('<entry>')\n if len(parts)>1:\n seg=parts[1]\n t=seg.split('<title>')[1].split('</title>')[0].strip().replace('\n',' ')\n title=t\n print(aid, title)\n except Exception as e:\n print('ERR', aid, e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?id_list=([2012.09289](/papers/2012.09289))'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(repr(data[:500]))\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","echo hi"]} “Projected Distribution Loss” is not a single standardized construct across arXiv literature. The expression and closely related variants are used for several technically distinct objects: a perceptual training objective for image enhancement based on projected feature distributions; a projected reconstruction term for intrinsic-dimension estimation in autoencoders; a camera-aware negative log-likelihood for probabilistic 3D regression; a forecast of the full conditional law of losses in actuarial modeling; a compound aggregate loss distribution for financial crises; and the estimation, reduction, or allocation of technical losses in electrical distribution networks (Delbracio et al., 2020, Orioua et al., 12 Sep 2025, Mohlin et al., 2023, Avanzi et al., 2024, Kapp et al., 2012, Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020). This suggests that the term is field-dependent: in some papers it denotes a training loss built from projected representations, while in others it denotes a projected or estimated distribution of losses, or physical loss in distribution systems.

1. Terminological scope and recurrent idea

Across the cited literature, the shared motif is the use of a projection, refinement, or conditional distribution to represent uncertainty or loss more faithfully than a simple point estimate. The technical object, however, changes by domain.

Domain	Meaning in the source literature	Source
Image enhancement	Perceptual loss using aggregated 1D-Wasserstein distances between projected CNN features	(Delbracio et al., 2020)
Intrinsic-dimension estimation	Projected reconstruction loss after removing one active latent dimension	(Orioua et al., 12 Sep 2025)
Probabilistic 3D regression	Negative log-likelihood of the projected Huber distribution	(Mohlin et al., 2023)
Actuarial modeling	Full conditional loss distribution $Y \mid \boldsymbol X=\boldsymbol x$ refined by a neural network	(Avanzi et al., 2024)
Financial crises	Aggregate GDP loss distribution from a Loss Distribution Approach	(Kapp et al., 2012)
Distribution grids	Technical loss estimation, reduction, or allocation in distribution networks	(Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020)

A common misconception is that “Projected Distribution Loss” names one canonical objective. In the cited work, it instead designates several non-equivalent constructions. Another recurring distinction is between a loss function used for training and a loss distribution used for forecasting or risk analysis. The former appears most clearly in image restoration, autoencoding, and 3D regression; the latter appears in actuarial, macro-financial, and power-system settings.

2. Image enhancement: Projected Distribution Loss as a perceptual objective

In image restoration, Projected Distribution Loss, or PDL, is a perceptual training objective that compares the distributions of deep features rather than their pointwise differences (Delbracio et al., 2020). The motivation is that pixel losses such as $L_1$ and $L_2$ encourage regression to the mean, and classic perceptual feature losses still compare features element-wise. Distribution-matching losses allow more flexibility, but the paper emphasizes that some are only approximate distribution matchers and adversarial objectives can produce hallucinations and are harder to optimize. PDL is positioned as an intermediate alternative: it keeps paired supervision from a reference image while relaxing strict pointwise alignment.

The formal change is to replace the direct feature norm by Wasserstein distances between projected feature distributions. With VGG feature extractor $\Phi$, feature maps $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$, projection vectors $w_j\in\mathbb{R}^m$, and projected features $\phi'_{i,j}(u)=w_j^T\phi_i(u)$, the loss is

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$

with $q=1$ and $p=1$ in all reported results (Delbracio et al., 2020). The implementation computes VGG16-conv4 features, projects them, sorts the projected samples, evaluates the 1D Wasserstein distance, sums across projections, and adds the pixel term. The paper notes that sorting is treated as fixed in backpropagation.

The paper evaluates PDL as a drop-in replacement for perceptual losses in denoising, single-image super-resolution, demosaicing, deblurring, and JPEG artifact removal. The architectures are SRN for denoising, demosaicing, and JPEG artifact removal, EDSR for super-resolution, and DsDeblur for deblurring. Training uses Adam, batch size $L_1$0, $L_1$1 iterations, VGG16-conv4 features, and DIV2K with synthetic degradations (Delbracio et al., 2020).

The reported results support the claim that the projected distribution view changes the perception–distortion trade-off. For denoising with $L_1$2, the best PDL row in Table 1 gives PSNR $L_1$3, MS-SSIM $L_1$4, LPIPS $L_1$5, NIQE $L_1$6, and FID $L_1$7. For $L_1$8 super-resolution, the best PDL result gives PSNR $L_1$9, MS-SSIM $L_2$0, LPIPS $L_2$1, NIQE $L_2$2, and FID $L_2$3. For deblurring, the best PDL result gives PSNR $L_2$4, MS-SSIM $L_2$5, LPIPS $L_2$6, NIQE $L_2$7, and FID $L_2$8 (Delbracio et al., 2020). Human preference studies reported pairwise preferences such as PDL vs no-perceptual $L_2$9, PDL vs $\Phi$0, PDL vs CTXDP $\Phi$1, and PDL vs CTXL2 $\Phi$2.

The ablation study shows that projection choice is not the dominant issue: Id, R2P, RPP, and RSP behave similarly, and the simplest channelwise marginal comparison is often as good as or better than more elaborate schemes. The paper also reports that PDL is only marginally slower than standard VGG perceptual loss, with training speeds of 0.57 steps/sec for PDL and 0.57 for $\Phi$3 perceptual loss (Delbracio et al., 2020).

3. Autoencoders and intrinsic dimension: projected reconstruction loss

In the Intrinsic Dimension Estimating Autoencoder (IDEA), the relevant notion is the projected reconstruction loss, also described as the projected loss or projected reconstruction term; the paper states that these are the same concept (Orioua et al., 12 Sep 2025). The purpose is not generic perceptual quality but the identification of the smallest effective latent dimension $\Phi$4 that still preserves reconstruction quality and should match the intrinsic dimension $\Phi$5.

IDEA uses a re-weighted double CancelOut layer. The first CancelOut layer, $\Phi$6, has trainable nonnegative weights and can suppress latent coordinates by pushing a weight to zero; the second, $\Phi$7, rescales the surviving latent features. The number of non-zero weights in $\Phi$8 is the effective latent size $\Phi$9 (Orioua et al., 12 Sep 2025). The projected term is introduced because a standard autoencoder can reconstruct well without revealing which latent dimensions are truly necessary.

The total training loss is written as a sum of four terms: reconstruction, projected reconstruction, L1 regularization, and orthogonality penalty. The projected reconstruction term is computed from $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$0, defined as “the output of the model after setting the last non-zero weight of the $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$1 layer to zero” (Orioua et al., 12 Sep 2025). Operationally, the model sees both the current reconstruction error $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$2 and the reconstruction error after removing one active dimension, $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$3. The paper characterizes elimination of a latent dimension as “a trade-off between the first two terms of the loss function.”

This projected look-ahead is the key mechanism. The paper states that it “anticipates the structure that the model will have if it decides to eliminate a dimension by setting the current $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$4 to zero” (Orioua et al., 12 Sep 2025). If the projected loss stays small, the dimension is likely redundant; if it rises sharply, the dimension is contributing real information. The orthogonality term $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$5 complements this by encouraging the latent space to become “uncorrelated.”

The empirical behavior is reported most explicitly on synthetic benchmarks. For Synthetic Legendre datasets with true dimensions $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$6, IDEA initialized with $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$7 correctly predicted $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$8. The paper reports that using $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$9 dimensions causes much larger test loss, while $w_j\in\mathbb{R}^m$0 does not improve reconstruction meaningfully; for dataset 1, $w_j\in\mathbb{R}^m$1, but $w_j\in\mathbb{R}^m$2 gives a test loss of $w_j\in\mathbb{R}^m$3, much worse than the $w_j\in\mathbb{R}^m$4 and $w_j\in\mathbb{R}^m$5 models around $w_j\in\mathbb{R}^m$6 (Orioua et al., 12 Sep 2025). The training dynamics are described in two phases: dimensions are progressively suppressed and each elimination causes a spike in the loss; then elimination stops once the trade-off is no longer favorable.

The same framework is applied to vertically resolved free-surface flow. In the simple case, IDEA predicts $w_j\in\mathbb{R}^m$7 total latent dimensions, including one dimension for water height, corresponding to a 2D manifold for the velocity profile. In the more complex case, it still predicts $w_j\in\mathbb{R}^m$8. The paper also reports that IDEA’s 2D velocity representation outperformed a truncated Legendre-moment projection that needed 4 moments in the simple case, and that in the more complex case the moment model needed at least 7 coefficients to reach comparable quality (Orioua et al., 12 Sep 2025).

4. Probabilistic 3D regression: projected Huber distribution and its loss

In monocular 3D regression, the relevant object is the negative log-likelihood of the projected Huber distribution (Mohlin et al., 2023). Here the “projected” language refers to the fact that image-plane coordinates and depth are coupled through perspective, and that a good probabilistic model should reflect this rather than predicting $w_j\in\mathbb{R}^m$9 naively.

The paper formulates seven desirable constraints for a camera-based 3D positional distribution: it should allow arbitrary variance in projected coordinates and depth independently; it should predict the projected location and depth normalized by focal length; it should assign zero probability to points behind the camera; it should assign positive probability to all physically visible points in the field of view; its negative log-likelihood should be convex in position; probability should vanish at infinity and at $\phi'_{i,j}(u)=w_j^T\phi_i(u)$0; and it should be continuous in position (Mohlin et al., 2023). The paper also proves that distributions of the form

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$1

cannot satisfy all these constraints simultaneously.

The proposed solution is the Projected Huber Distribution, composed of a projected-coordinate term conditioned on depth and a depth term with support only on $\phi'_{i,j}(u)=w_j^T\phi_i(u)$2. The projected part uses a Huber-style radial penalty, while the depth term uses

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$3

The combined distribution is $\phi'_{i,j}(u)=w_j^T\phi_i(u)$4 (Mohlin et al., 2023).

The training objective is the negative log-likelihood of this combined distribution. After a basis change to normalized coordinates,

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$5

the loss becomes

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$6

with constant terms omitted in implementation (Mohlin et al., 2023). Parameter remapping is used so that $\phi'_{i,j}(u)=w_j^T\phi_i(u)$7, $\phi'_{i,j}(u)=w_j^T\phi_i(u)$8, and $\phi'_{i,j}(u)=w_j^T\phi_i(u)$9.

This construction differs from standard MAE/MSE regression, Gaussian regression, plain Huber loss, and independent projected-and-depth probabilistic models. The paper’s emphasis is that the contribution is not just a new loss but a distributional form whose log-likelihood is tailored to camera geometry and downstream fusion (Mohlin et al., 2023). The mode used for point prediction is

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$0

Empirically, the experiments are synthetic and test focal-length variation, scale ambiguity, occlusion, and combinations of those factors. The reported conclusions are that predicted variances are well correlated with empirical squared errors; scale ambiguity raises depth uncertainty while projected uncertainty stays low; occlusion raises both projected and depth uncertainty; and the mode often outperforms MAE baselines on $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$1 and $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$2 (Mohlin et al., 2023).

5. Projected loss distributions in actuarial and macro-financial modeling

A different usage appears in actuarial science, where the object is the full conditional law of a loss variable $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$3 given covariates $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$4. The Distributional Refinement Network (DRN) defines this as the projected loss distribution, its tail behavior, and derived risk measures such as VaR (Avanzi et al., 2024). The emphasis is not on a training loss over projections, but on distributional forecasting.

The DRN begins from a standard parametric distributional regression model, typically a GLM, then refines the baseline distribution rather than replacing it. The baseline density is transformed into partitioned probability masses over intervals $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$5, with

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$6

The neural network takes the original covariates and the baseline bin masses as inputs, produces raw outputs $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$7, and refines the baseline masses multiplicatively through a softmax-type adjustment (Avanzi et al., 2024). The paper stresses that this preserves baseline structure and interpretability, because the final forecast can be decomposed into a transparent baseline component and a learned correction.

Training can be done with negative log-likelihood or the paper’s preferred joint binary cross-entropy (JBCE) objective. The regularized objective adds a KL term that keeps the DRN close to the baseline “prior,” a roughness penalty that reduces jagged density fluctuations, and a mean penalty that prevents unnecessary drift from the baseline mean when the GLM already captures the center well (Avanzi et al., 2024). On synthetic data, DRN improves test metrics from GLM values of NLL $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$8, CRPS $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$9, RMSE $q=1$0, and 90% quantile loss $q=1$1 to DRN values of NLL $q=1$2, CRPS $q=1$3, RMSE $q=1$4, and 90% quantile loss $q=1$5. On freMPL1 claim severity data, the paper reports GLM values of NLL $q=1$6, CRPS $q=1$7, RMSE $q=1$8, and 90% quantile loss $q=1$9, versus DRN values of NLL $p=1$0, CRPS $p=1$1, RMSE $p=1$2, and 90% quantile loss $p=1$3 (Avanzi et al., 2024).

The macro-financial literature uses a related but distinct framework. “Real Output Costs of Financial Crises: A Loss Distribution Approach” models crisis risk by separating frequency and severity and combining them into a global aggregate loss probability density function over five-year periods (Kapp et al., 2012). The aggregate loss is

$p=1$4

where $p=1$5 is the number of crisis events and $p=1$6 is the severity of event $p=1$7. Frequency is modeled with Poisson and Negative Binomial distributions, with the Negative Binomial preferred because the data are overdispersed; severity is fitted with Gamma, Exponential, GEV, Generalized Pareto, Log-normal, and Weibull, with Weibull used as the benchmark (Kapp et al., 2012). The reported specification uses 500,000 Monte Carlo simulations.

The resulting aggregate loss PDFs are “markedly skewed to the right.” The key headline result is that one-percent-probability global crises every five years lead to losses between $p=1$8 and $p=1$9 of world GDP, corresponding to the 99th percentile under different output-gap specifications. The 99.9th percentile reaches $L_1$00 to $L_1$01 of world GDP, while mean five-year world-GDP losses range from $L_1$02 to $L_1$03 and median losses from $L_1$04 to $L_1$05 (Kapp et al., 2012). This use of “loss distribution” is actuarial in the strict sense: it is a projected distribution of aggregate downside outcomes, not a neural-network training criterion.

6. Distribution grids: technical loss estimation, reduction, and allocation

In electrical distribution systems, the language of distribution loss refers to physical energy dissipated within the network, primarily through Joule heating, and the technical questions are how to estimate, reduce, or allocate it (Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020). This is again different from perceptual or probabilistic losses.

A first strand is voltage-only technical loss estimation. The method called voss (“Voltage to Technical Loss”) estimates the technical loss fraction of a line segment from endpoint voltage magnitudes without directly measuring currents or power flows (Bariya et al., 26 Jun 2025). For a single line segment,

$L_1$06

For a line with intervening loads, the corrected estimator is

$L_1$07

with

$L_1$08

The method is evaluated on the IEEE 13-node and 34-node feeders in OpenDSS. For line 816–822, phase A, the single-segment estimate is $L_1$09, the correction factor is $L_1$10, the multi-segment estimate is $L_1$11, and the true loss is $L_1$12, reducing the absolute error from about $L_1$13 to $L_1$14 (Bariya et al., 26 Jun 2025). The paper explicitly frames this as a practical estimator for LMIC distribution grids where conventional power-flow instrumentation is unavailable or impractical.

A second strand is coordinated loss reduction by co-optimizing Conservation Voltage Reduction (CVR) and Network Topology Reconfiguration (NTR). The objective is to minimize total active power losses over the scheduling horizon,

$L_1$15

using a mixed-integer conic program with AC branch-flow constraints, voltage-dependent ZIP loads, and radiality constraints (Fatima et al., 11 May 2026). CVR lowers nodal voltages to reduce voltage-sensitive demand and hence $L_1$16 losses, while NTR changes the feeder’s radial switching pattern to route power through lower-impedance paths and flatten branch loading. The paper argues that the interaction is multiplicative rather than additive: CVR reduces current magnitude, which enlarges the feasible region for reconfiguration, and NTR then moves those reduced currents onto lower-loss paths.

The reported numerical outcomes show that the coordinated CEDNTR case gives the lowest losses among all cases. On the IEEE 123-bus feeder, SDN has about $L_1$17 kW losses at $L_1$18 MW system load, SDNTR about $L_1$19 kW with $L_1$20 reduction, CEDN about $L_1$21 kW with $L_1$22 reduction, and CEDNTR about $L_1$23 kW with $L_1$24 reduction; the conclusion section also states a coordinated loss level of about $L_1$25 kW (Fatima et al., 11 May 2026). Maximum line loading decreases from about $L_1$26 in SDN to about $L_1$27 in CEDNTR, roughly a $L_1$28 reduction in maximum loading. The framework is tested under hourly load variation, three load compositions—constant impedance, constant current, and constant power—and with and without DERs including PV, battery energy storage, and dispatchable generation.

A third strand is loss allocation in joint transmission and distribution peer-to-peer markets. Here losses are modeled as an explicit market product: each bilateral trade $L_1$29 is accompanied by an injection $L_1$30 and a loss quantity $L_1$31, and a market operator allocates line losses through a loss-allocation matrix $L_1$32 (Moret et al., 2020). The DSO uses a linear AC approximation with distribution losses

$L_1$33

while the TSO uses

$L_1$34

The paper analyzes socialization, individual, capacity-weighted individual, and mixed allocation policies, concluding that socialization is fairer at DSO level, while individual allocation is more efficient but potentially inequitable in radial distribution grids (Moret et al., 2020).

Taken together, these power-system papers show that “distribution loss” can denote a measurable physical quantity, an optimization target, or an allocable market object. This suggests that in infrastructure research the phrase is anchored in network physics and operational planning rather than in representation learning or probabilistic perception.\ntry:\n print(urllib.request.urlopen(url, timeout=20).read().decode()[:1000])\nexcept Exception as e:\n print(\"ERR\", e)\nPY"]' to=shell 彩神争霸如何 code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?search_query=all:%22Projected+Distribution+Loss%22&start=0&max_results=5'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(data[:1200])\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?search_query=all:%22Projected+Distribution+Loss%22&start=0&max_results=5'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(data[:1200])\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nids=['(Delbracio et al., 2020)','(Orioua et al., 12 Sep 2025)','(Mohlin et al., 2023)','(Avanzi et al., 2024)','(Fatima et al., 11 May 2026)']\nfor aid in ids:\n url=f'http://export.arxiv.org/api/query?id_list={aid}'\n try:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n title=''\n if '<title>' in data:\n parts=data.split('<entry>')\n if len(parts)>1:\n seg=parts[1]\n t=seg.split('<title>')[1].split('</title>')[0].strip().replace('\n',' ')\n title=t\n print(aid, title)\n except Exception as e:\n print('ERR', aid, e)\nPY"]} to=shell code={"command":["bash","-lc","python - <<'PY'\nimport urllib.request\nurl='http://export.arxiv.org/api/query?id_list=([2012.09289](/papers/2012.09289))'\ntry:\n data=urllib.request.urlopen(url, timeout=20).read().decode()\n print(repr(data[:500]))\nexcept Exception as e:\n print('ERR', e)\nPY"]} to=shell code={"command":["bash","-lc","echo hi"]} “Projected Distribution Loss” is not a single standardized construct across arXiv literature. The expression and closely related variants are used for several technically distinct objects: a perceptual training objective for image enhancement based on projected feature distributions; a projected reconstruction term for intrinsic-dimension estimation in autoencoders; a camera-aware negative log-likelihood for probabilistic 3D regression; a forecast of the full conditional law of losses in actuarial modeling; a compound aggregate loss distribution for financial crises; and the estimation, reduction, or allocation of technical losses in electrical distribution networks (Delbracio et al., 2020, Orioua et al., 12 Sep 2025, Mohlin et al., 2023, Avanzi et al., 2024, Kapp et al., 2012, Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020). This suggests that the term is field-dependent: in some papers it denotes a training loss built from projected representations, while in others it denotes a projected or estimated distribution of losses, or physical loss in distribution systems.

1. Terminological scope and recurrent idea

Across the cited literature, the shared motif is the use of a projection, refinement, or conditional distribution to represent uncertainty or loss more faithfully than a simple point estimate. The technical object, however, changes by domain.

Domain	Meaning in the source literature	Source
Image enhancement	Perceptual loss using aggregated 1D-Wasserstein distances between projected CNN features	(Delbracio et al., 2020)
Intrinsic-dimension estimation	Projected reconstruction loss after removing one active latent dimension	(Orioua et al., 12 Sep 2025)
Probabilistic 3D regression	Negative log-likelihood of the projected Huber distribution	(Mohlin et al., 2023)
Actuarial modeling	Full conditional loss distribution $Y \mid \boldsymbol X=\boldsymbol x$ refined by a neural network	(Avanzi et al., 2024)
Financial crises	Aggregate GDP loss distribution from a Loss Distribution Approach	(Kapp et al., 2012)
Distribution grids	Technical loss estimation, reduction, or allocation in distribution networks	(Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020)

A common misconception is that “Projected Distribution Loss” names one canonical objective. In the cited work, it instead designates several non-equivalent constructions. Another recurring distinction is between a loss function used for training and a loss distribution used for forecasting or risk analysis. The former appears most clearly in image restoration, autoencoding, and 3D regression; the latter appears in actuarial, macro-financial, and power-system settings.

2. Image enhancement: Projected Distribution Loss as a perceptual objective

In image restoration, Projected Distribution Loss, or PDL, is a perceptual training objective that compares the distributions of deep features rather than their pointwise differences (Delbracio et al., 2020). The motivation is that pixel losses such as $L_1$ and $L_2$ encourage regression to the mean, and classic perceptual feature losses still compare features element-wise. Distribution-matching losses allow more flexibility, but the paper emphasizes that some are only approximate distribution matchers and adversarial objectives can produce hallucinations and are harder to optimize. PDL is positioned as an intermediate alternative: it keeps paired supervision from a reference image while relaxing strict pointwise alignment.

The formal change is to replace the direct feature norm by Wasserstein distances between projected feature distributions. With VGG feature extractor $\Phi$, feature maps $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$, projection vectors $w_j\in\mathbb{R}^m$, and projected features $\phi'_{i,j}(u)=w_j^T\phi_i(u)$, the loss is

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$

with $q=1$ and $p=1$ in all reported results (Delbracio et al., 2020). The implementation computes VGG16-conv4 features, projects them, sorts the projected samples, evaluates the 1D Wasserstein distance, sums across projections, and adds the pixel term. The paper notes that sorting is treated as fixed in backpropagation.

The paper evaluates PDL as a drop-in replacement for perceptual losses in denoising, single-image super-resolution, demosaicing, deblurring, and JPEG artifact removal. The architectures are SRN for denoising, demosaicing, and JPEG artifact removal, EDSR for super-resolution, and DsDeblur for deblurring. Training uses Adam, batch size $L_1$0, $L_1$1 iterations, VGG16-conv4 features, and DIV2K with synthetic degradations (Delbracio et al., 2020).

The reported results support the claim that the projected distribution view changes the perception–distortion trade-off. For denoising with $L_1$2, the best PDL row in Table 1 gives PSNR $L_1$3, MS-SSIM $L_1$4, LPIPS $L_1$5, NIQE $L_1$6, and FID $L_1$7. For $L_1$8 super-resolution, the best PDL result gives PSNR $L_1$9, MS-SSIM $L_2$0, LPIPS $L_2$1, NIQE $L_2$2, and FID $L_2$3. For deblurring, the best PDL result gives PSNR $L_2$4, MS-SSIM $L_2$5, LPIPS $L_2$6, NIQE $L_2$7, and FID $L_2$8 (Delbracio et al., 2020). Human preference studies reported pairwise preferences such as PDL vs no-perceptual $L_2$9, PDL vs $\Phi$0, PDL vs CTXDP $\Phi$1, and PDL vs CTXL2 $\Phi$2.

The ablation study shows that projection choice is not the dominant issue: Id, R2P, RPP, and RSP behave similarly, and the simplest channelwise marginal comparison is often as good as or better than more elaborate schemes. The paper also reports that PDL is only marginally slower than standard VGG perceptual loss, with training speeds of 0.57 steps/sec for PDL and 0.57 for $\Phi$3 perceptual loss (Delbracio et al., 2020).

3. Autoencoders and intrinsic dimension: projected reconstruction loss

In the Intrinsic Dimension Estimating Autoencoder (IDEA), the relevant notion is the projected reconstruction loss, also described as the projected loss or projected reconstruction term; the paper states that these are the same concept (Orioua et al., 12 Sep 2025). The purpose is not generic perceptual quality but the identification of the smallest effective latent dimension $\Phi$4 that still preserves reconstruction quality and should match the intrinsic dimension $\Phi$5.

IDEA uses a re-weighted double CancelOut layer. The first CancelOut layer, $\Phi$6, has trainable nonnegative weights and can suppress latent coordinates by pushing a weight to zero; the second, $\Phi$7, rescales the surviving latent features. The number of non-zero weights in $\Phi$8 is the effective latent size $\Phi$9 (Orioua et al., 12 Sep 2025). The projected term is introduced because a standard autoencoder can reconstruct well without revealing which latent dimensions are truly necessary.

The total training loss is written as a sum of four terms: reconstruction, projected reconstruction, L1 regularization, and orthogonality penalty. The projected reconstruction term is computed from $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$0, defined as “the output of the model after setting the last non-zero weight of the $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$1 layer to zero” (Orioua et al., 12 Sep 2025). Operationally, the model sees both the current reconstruction error $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$2 and the reconstruction error after removing one active dimension, $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$3. The paper characterizes elimination of a latent dimension as “a trade-off between the first two terms of the loss function.”

This projected look-ahead is the key mechanism. The paper states that it “anticipates the structure that the model will have if it decides to eliminate a dimension by setting the current $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$4 to zero” (Orioua et al., 12 Sep 2025). If the projected loss stays small, the dimension is likely redundant; if it rises sharply, the dimension is contributing real information. The orthogonality term $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$5 complements this by encouraging the latent space to become “uncorrelated.”

The empirical behavior is reported most explicitly on synthetic benchmarks. For Synthetic Legendre datasets with true dimensions $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$6, IDEA initialized with $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$7 correctly predicted $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$8. The paper reports that using $\Phi(u), \Phi(v)\in\mathbb{R}^{n\times m}$9 dimensions causes much larger test loss, while $w_j\in\mathbb{R}^m$0 does not improve reconstruction meaningfully; for dataset 1, $w_j\in\mathbb{R}^m$1, but $w_j\in\mathbb{R}^m$2 gives a test loss of $w_j\in\mathbb{R}^m$3, much worse than the $w_j\in\mathbb{R}^m$4 and $w_j\in\mathbb{R}^m$5 models around $w_j\in\mathbb{R}^m$6 (Orioua et al., 12 Sep 2025). The training dynamics are described in two phases: dimensions are progressively suppressed and each elimination causes a spike in the loss; then elimination stops once the trade-off is no longer favorable.

The same framework is applied to vertically resolved free-surface flow. In the simple case, IDEA predicts $w_j\in\mathbb{R}^m$7 total latent dimensions, including one dimension for water height, corresponding to a 2D manifold for the velocity profile. In the more complex case, it still predicts $w_j\in\mathbb{R}^m$8. The paper also reports that IDEA’s 2D velocity representation outperformed a truncated Legendre-moment projection that needed 4 moments in the simple case, and that in the more complex case the moment model needed at least 7 coefficients to reach comparable quality (Orioua et al., 12 Sep 2025).

4. Probabilistic 3D regression: projected Huber distribution and its loss

In monocular 3D regression, the relevant object is the negative log-likelihood of the projected Huber distribution (Mohlin et al., 2023). Here the “projected” language refers to the fact that image-plane coordinates and depth are coupled through perspective, and that a good probabilistic model should reflect this rather than predicting $w_j\in\mathbb{R}^m$9 naively.

The paper formulates seven desirable constraints for a camera-based 3D positional distribution: it should allow arbitrary variance in projected coordinates and depth independently; it should predict the projected location and depth normalized by focal length; it should assign zero probability to points behind the camera; it should assign positive probability to all physically visible points in the field of view; its negative log-likelihood should be convex in position; probability should vanish at infinity and at $\phi'_{i,j}(u)=w_j^T\phi_i(u)$0; and it should be continuous in position (Mohlin et al., 2023). The paper also proves that distributions of the form

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$1

cannot satisfy all these constraints simultaneously.

The proposed solution is the Projected Huber Distribution, composed of a projected-coordinate term conditioned on depth and a depth term with support only on $\phi'_{i,j}(u)=w_j^T\phi_i(u)$2. The projected part uses a Huber-style radial penalty, while the depth term uses

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$3

The combined distribution is $\phi'_{i,j}(u)=w_j^T\phi_i(u)$4 (Mohlin et al., 2023).

The training objective is the negative log-likelihood of this combined distribution. After a basis change to normalized coordinates,

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$5

the loss becomes

$\phi'_{i,j}(u)=w_j^T\phi_i(u)$6

with constant terms omitted in implementation (Mohlin et al., 2023). Parameter remapping is used so that $\phi'_{i,j}(u)=w_j^T\phi_i(u)$7, $\phi'_{i,j}(u)=w_j^T\phi_i(u)$8, and $\phi'_{i,j}(u)=w_j^T\phi_i(u)$9.

This construction differs from standard MAE/MSE regression, Gaussian regression, plain Huber loss, and independent projected-and-depth probabilistic models. The paper’s emphasis is that the contribution is not just a new loss but a distributional form whose log-likelihood is tailored to camera geometry and downstream fusion (Mohlin et al., 2023). The mode used for point prediction is

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$0

Empirically, the experiments are synthetic and test focal-length variation, scale ambiguity, occlusion, and combinations of those factors. The reported conclusions are that predicted variances are well correlated with empirical squared errors; scale ambiguity raises depth uncertainty while projected uncertainty stays low; occlusion raises both projected and depth uncertainty; and the mode often outperforms MAE baselines on $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$1 and $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$2 (Mohlin et al., 2023).

5. Projected loss distributions in actuarial and macro-financial modeling

A different usage appears in actuarial science, where the object is the full conditional law of a loss variable $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$3 given covariates $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$4. The Distributional Refinement Network (DRN) defines this as the projected loss distribution, its tail behavior, and derived risk measures such as VaR (Avanzi et al., 2024). The emphasis is not on a training loss over projections, but on distributional forecasting.

The DRN begins from a standard parametric distributional regression model, typically a GLM, then refines the baseline distribution rather than replacing it. The baseline density is transformed into partitioned probability masses over intervals $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$5, with

$$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$6

The neural network takes the original covariates and the baseline bin masses as inputs, produces raw outputs $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$7, and refines the baseline masses multiplicatively through a softmax-type adjustment (Avanzi et al., 2024). The paper stresses that this preserves baseline structure and interpretability, because the final forecast can be decomposed into a transparent baseline component and a learned correction.

Training can be done with negative log-likelihood or the paper’s preferred joint binary cross-entropy (JBCE) objective. The regularized objective adds a KL term that keeps the DRN close to the baseline “prior,” a roughness penalty that reduces jagged density fluctuations, and a mean penalty that prevents unnecessary drift from the baseline mean when the GLM already captures the center well (Avanzi et al., 2024). On synthetic data, DRN improves test metrics from GLM values of NLL $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$8, CRPS $$L_{pdl}(u, v) = \|u - v\|_q + \lambda \sum_{j=1}^{m'} W_p \left(\phi'_j(u), \phi'_j(v) \right),$$9, RMSE $q=1$0, and 90% quantile loss $q=1$1 to DRN values of NLL $q=1$2, CRPS $q=1$3, RMSE $q=1$4, and 90% quantile loss $q=1$5. On freMPL1 claim severity data, the paper reports GLM values of NLL $q=1$6, CRPS $q=1$7, RMSE $q=1$8, and 90% quantile loss $q=1$9, versus DRN values of NLL $p=1$0, CRPS $p=1$1, RMSE $p=1$2, and 90% quantile loss $p=1$3 (Avanzi et al., 2024).

The macro-financial literature uses a related but distinct framework. “Real Output Costs of Financial Crises: A Loss Distribution Approach” models crisis risk by separating frequency and severity and combining them into a global aggregate loss probability density function over five-year periods (Kapp et al., 2012). The aggregate loss is

$p=1$4

where $p=1$5 is the number of crisis events and $p=1$6 is the severity of event $p=1$7. Frequency is modeled with Poisson and Negative Binomial distributions, with the Negative Binomial preferred because the data are overdispersed; severity is fitted with Gamma, Exponential, GEV, Generalized Pareto, Log-normal, and Weibull, with Weibull used as the benchmark (Kapp et al., 2012). The reported specification uses 500,000 Monte Carlo simulations.

The resulting aggregate loss PDFs are “markedly skewed to the right.” The key headline result is that one-percent-probability global crises every five years lead to losses between $p=1$8 and $p=1$9 of world GDP, corresponding to the 99th percentile under different output-gap specifications. The 99.9th percentile reaches $L_1$00 to $L_1$01 of world GDP, while mean five-year world-GDP losses range from $L_1$02 to $L_1$03 and median losses from $L_1$04 to $L_1$05 (Kapp et al., 2012). This use of “loss distribution” is actuarial in the strict sense: it is a projected distribution of aggregate downside outcomes, not a neural-network training criterion.

6. Distribution grids: technical loss estimation, reduction, and allocation

In electrical distribution systems, the language of distribution loss refers to physical energy dissipated within the network, primarily through Joule heating, and the technical questions are how to estimate, reduce, or allocate it (Bariya et al., 26 Jun 2025, Fatima et al., 11 May 2026, Moret et al., 2020). This is again different from perceptual or probabilistic losses.

A first strand is voltage-only technical loss estimation. The method called voss (“Voltage to Technical Loss”) estimates the technical loss fraction of a line segment from endpoint voltage magnitudes without directly measuring currents or power flows (Bariya et al., 26 Jun 2025). For a single line segment,

$L_1$06

For a line with intervening loads, the corrected estimator is

$L_1$07

with

$L_1$08

The method is evaluated on the IEEE 13-node and 34-node feeders in OpenDSS. For line 816–822, phase A, the single-segment estimate is $L_1$09, the correction factor is $L_1$10, the multi-segment estimate is $L_1$11, and the true loss is $L_1$12, reducing the absolute error from about $L_1$13 to $L_1$14 (Bariya et al., 26 Jun 2025). The paper explicitly frames this as a practical estimator for LMIC distribution grids where conventional power-flow instrumentation is unavailable or impractical.

A second strand is coordinated loss reduction by co-optimizing Conservation Voltage Reduction (CVR) and Network Topology Reconfiguration (NTR). The objective is to minimize total active power losses over the scheduling horizon,

$L_1$15

using a mixed-integer conic program with AC branch-flow constraints, voltage-dependent ZIP loads, and radiality constraints (Fatima et al., 11 May 2026). CVR lowers nodal voltages to reduce voltage-sensitive demand and hence $L_1$16 losses, while NTR changes the feeder’s radial switching pattern to route power through lower-impedance paths and flatten branch loading. The paper argues that the interaction is multiplicative rather than additive: CVR reduces current magnitude, which enlarges the feasible region for reconfiguration, and NTR then moves those reduced currents onto lower-loss paths.

The reported numerical outcomes show that the coordinated CEDNTR case gives the lowest losses among all cases. On the IEEE 123-bus feeder, SDN has about $L_1$17 kW losses at $L_1$18 MW system load, SDNTR about $L_1$19 kW with $L_1$20 reduction, CEDN about $L_1$21 kW with $L_1$22 reduction, and CEDNTR about $L_1$23 kW with $L_1$24 reduction; the conclusion section also states a coordinated loss level of about $L_1$25 kW (Fatima et al., 11 May 2026). Maximum line loading decreases from about $L_1$26 in SDN to about $L_1$27 in CEDNTR, roughly a $L_1$28 reduction in maximum loading. The framework is tested under hourly load variation, three load compositions—constant impedance, constant current, and constant power—and with and without DERs including PV, battery energy storage, and dispatchable generation.

A third strand is loss allocation in joint transmission and distribution peer-to-peer markets. Here losses are modeled as an explicit market product: each bilateral trade $L_1$29 is accompanied by an injection $L_1$30 and a loss quantity $L_1$31, and a market operator allocates line losses through a loss-allocation matrix $L_1$32 (Moret et al., 2020). The DSO uses a linear AC approximation with distribution losses

$L_1$33

while the TSO uses

$L_1$34

The paper analyzes socialization, individual, capacity-weighted individual, and mixed allocation policies, concluding that socialization is fairer at DSO level, while individual allocation is more efficient but potentially inequitable in radial distribution grids (Moret et al., 2020).

Taken together, these power-system papers show that “distribution loss” can denote a measurable physical quantity, an optimization target, or an allocable market object. This suggests that in infrastructure research the phrase is anchored in network physics and operational planning rather than in representation learning or probabilistic perception.