Target Audience
- Readers interested in AI-driven protein generation and its applications in drug discovery and pharmaceutical development.
- Researchers and practitioners who want to understand the differences between models such as EvoDiff, RFdiffusion, and ESM3, and how they fit into the history of protein generation research.
- Anyone interested in recent advances in generative AI models that integrate protein sequences, structures, functions, and molecular interactions.
Summary
- EvoDiff, introduced in 2023, is a discrete diffusion model for protein generation that directly leverages evolutionary sequence data.
- At a time when structure-based protein generation was gaining momentum, EvoDiff proposed a sequence-first approach that explores protein sequence space without explicitly generating three-dimensional structures.
- This research helped establish sequence diffusion as an important direction, influencing subsequent models such as DPLM and demonstrating the value of large-scale evolutionary sequence data.
- Meanwhile, the frontier of protein generation research in 2025–2026 has shifted toward multimodal and all-atom generative models that integrate sequences, structures, functions, and molecular interactions.
- The historical significance of EvoDiff lies not in proving that structural information is unnecessary, but in establishing sequence-first generation as an independent and valuable research direction.
1. Before EvoDiff: Learning Proteins as a Language
To understand EvoDiff, it is useful to first examine how research on learning from protein sequences has evolved.
Amino acid sequences contain various constraints that have been preserved throughout evolution.
This led researchers to explore whether statistical patterns learned from large collections of naturally occurring protein sequences could capture information about protein structures and functions.
Around 2017–2019: From Statistical Models of Protein Families to Deep Generative Models
One of the important early studies in this area was DeepSequence.
DeepSequence used a Variational Autoencoder (VAE) to learn the sequence distribution represented in a Multiple Sequence Alignment (MSA), modeling statistical features of sequences that had been shaped by natural evolution.
Although its primary objective was to predict the effects of mutations, the study was important because it demonstrated the idea of learning evolutionary sequence distributions using a generative model. [1]
Subsequent research, including UniRep, showed that self-supervised learning on large collections of protein sequences could produce useful representations related to protein structure and function, even when learning exclusively from sequences. [2]
Together, these studies established the foundation for large-scale learning directly from protein sequences.
Around 2020–2023: Direct Generation with Protein Language Models
Protein Language Models (PLMs) based on Transformer architectures subsequently became an important research direction.
Representative examples include ProGen, ProGen2, and ProtGPT2.
These models treat amino acid sequences similarly to natural language and generate new protein sequences by predicting the next amino acid.
A typical autoregressive model learns the following probability distribution:
$$ P(x_1,\ldots,x_L)
\prod_{i=1}^{L} P(x_i \mid x_{<i}) $$
Here, $x_i$ represents the amino acid at position $i$ in the sequence.
In particular, the ProGen study demonstrated that a model with 1.2 billion parameters, trained on approximately 280 million protein sequences, could generate artificial proteins that were experimentally confirmed to be functional. [3]
This result demonstrated that explicitly modeling three-dimensional protein structures is not always necessary to build useful protein generative models.
Therefore, the idea that "sequence is enough" did not originate with EvoDiff.
What made EvoDiff novel was its combination of three elements:
- Sequence language models
- Diffusion models
- Evolutionary-scale sequence data
2. A Parallel Development: The Rapid Rise of Structure Generation in 2022–2023
Alongside research on directly generating protein sequences, protein design based on three-dimensional structures was also advancing rapidly.
ProteinMPNN played an important role in this development.
ProteinMPNN is an inverse folding model that designs amino acid sequences likely to adopt a given protein backbone structure.
In other words, it made the following design direction highly practical:
3D Backbone → Amino Acid Sequence
Then, in 2023, RFdiffusion and Chroma emerged.
RFdiffusion and Chroma
RFdiffusion uses structural representations from RoseTTAFold and applies diffusion to three-dimensional coordinates to generate new protein backbones.
This research led to significant advances in tasks such as:
- De novo backbone generation
- Motif scaffolding
- Binder design
Chroma, which appeared around the same time, enabled large-scale conditional generation of three-dimensional protein structures. Its study was published in Nature in November 2023. [4]
A typical protein generation pipeline during this period looked like this:
Random Noise → 3D Backbone → ProteinMPNN → Amino Acid Sequence
First, a diffusion model generates a three-dimensional backbone. Another model then designs an amino acid sequence compatible with that structure.
In other words, structure-first approaches, which place three-dimensional structures at the center of protein generation, were gaining momentum.
It was against this research backdrop that EvoDiff emerged.
3. The Emergence of EvoDiff: Generating Evolutionary Sequences Directly
EvoDiff was introduced by Alamdari and colleagues in their study, "Protein generation with evolutionary diffusion: sequence is all you need."
The original bioRxiv preprint was released on September 12, 2023. [5]
The central question behind EvoDiff is remarkably simple:
Why do we need to generate a three-dimensional structure first when designing proteins?
Nature contains an enormous collection of protein sequences accumulated through evolution.
If a generative model could learn directly from those sequences, perhaps it could generate useful proteins without explicitly constructing their three-dimensional structures.
This is the fundamental idea behind EvoDiff.
EvoDiff trains a discrete diffusion model on amino acid sequences using approximately 42 million protein sequences derived from UniRef50.
This model is called EvoDiff-Seq.
The authors also proposed EvoDiff-MSA, which applies diffusion directly to multiple sequence alignments. [5]
The significance of EvoDiff extends beyond the introduction of another generative algorithm.
Its central contribution was shifting the generation space from three-dimensional structures to evolutionary sequences themselves.
4. EvoDiff's First Major Contribution: Leveraging the Scale of Evolutionary Sequence Data
One of the most important points emphasized by EvoDiff is the difference in scale between available protein sequence data and experimentally determined three-dimensional structures.
At the time, the number of experimentally solved protein structures was on the order of hundreds of thousands.
By contrast, the number of protein sequences collected from nature was orders of magnitude larger.
This asymmetry was a central motivation for EvoDiff. [6]
Structure-First vs. Sequence-First
Traditional structure-based generative models conceptually learn a distribution such as:
$$ P(\text{structure}) $$
Their training data therefore primarily consists of protein structures.
In contrast, EvoDiff directly learns:
$$ P(\text{sequence}) $$
By modeling evolutionary sequences themselves, EvoDiff can leverage a much larger pool of available training data.
This idea is closely related to the philosophy behind modern protein foundation models.
Rather than restricting protein generation to structural datasets, EvoDiff places the vast amount of sequence information preserved by natural evolution at the center of the modeling process.
This is one of its most important contributions.
5. EvoDiff's Second Major Contribution: Generation Does Not Have to Proceed from Left to Right
Another important feature of EvoDiff is its generation process, which differs fundamentally from autoregressive models.
In autoregressive models such as ProGen, amino acid sequences are generally generated sequentially from left to right:
$$ 1 \rightarrow 2 \rightarrow 3 \rightarrow \cdots \rightarrow L $$
EvoDiff instead learns to gradually reconstruct amino acid sequences from states in which some residues have been masked or corrupted with noise.
Conceptually, generation can look like this:
???A????G???
↓
MKTAV...G...The key advantage is that arbitrary regions of a sequence can become generation targets.
Flexible Conditional Generation
This property allows EvoDiff to naturally support tasks such as:
- Unconditional generation: Generating new protein sequences without specifying additional conditions.
- Sequence inpainting: Keeping part of a sequence fixed while generating missing regions.
- Motif scaffolding: Preserving functionally important motifs while generating surrounding sequences.
- Fixing the N- and C-termini: Maintaining both terminal regions while generating the sequence between them.
- MSA-conditioned generation: Generating sequences using information from evolutionarily related sequences.
This flexibility is an important advantage of applying diffusion models to protein sequences.
It also represents a feature that connects naturally to subsequent research on diffusion language models.
6. EvoDiff's Third Major Contribution: Handling Proteins That Are Challenging for Structure-Based Models
Another important aspect of EvoDiff is that it can model proteins and regions that do not possess a single stable three-dimensional structure.
Structure-generating models such as RFdiffusion are particularly effective for designing proteins with well-defined three-dimensional structures.
However, many biologically important proteins and regions do not function through a single stable conformation.
Examples include:
- Intrinsically disordered proteins and regions (IDPs / IDRs)
- Flexible linkers
- Targeting peptides
- Signal sequences
For these regions, flexibility and structural disorder can be directly related to biological function.
Because EvoDiff generates amino acid sequences directly, it can, in principle, model these regions as ordinary sequence-generation targets.
The paper positions this approach as sequence-first design that goes beyond the traditional emphasis on the structure–function paradigm. [5]
This remains one of EvoDiff's most distinctive contributions, even when viewed from the perspective of current research.
7. Was "Sequence Is All You Need" Correct?
The title of the EvoDiff paper makes a strong claim: "sequence is all you need."
Looking back from 2026, it is useful to separate this claim into two interpretations.
Powerful Generative Models Can Be Trained Using Sequence Data Alone
Subsequent research has supported this claim.
The ability to learn diverse biological constraints and features from evolutionary sequence data alone is one of the central achievements of protein language model research.
EvoDiff demonstrated one effective approach to this problem through sequence diffusion.
Protein Design Does Not Require Information Beyond Sequence
For this stronger interpretation, however, the direction of current research suggests a different conclusion.
Subsequent models have increasingly integrated structural and functional information while retaining the advantages of sequence-based learning.
The overall progression can be summarized as:
$$ \text{Sequence} \rightarrow \text{Sequence + Structure} \rightarrow \text{Sequence + Structure + Function + Interactions} $$
Therefore, it is more appropriate to interpret EvoDiff's title as a challenge to the assumption that three-dimensional structural data must always be at the center of protein generation.
It should not be interpreted as proof that structural information is unnecessary for protein design.
8. A Direct Successor to EvoDiff: DPLM
An important follow-up study in the evolution of EvoDiff's ideas is DPLM, or Diffusion Protein Language Model.
DPLM was proposed by Wang and colleagues and presented at ICML 2024. [7]
It extends sequence diffusion beyond protein generation and explores its use in representation learning.
From a Generative Model to a Foundation Model
The importance of DPLM lies in positioning diffusion models not merely as sequence generators, but as general-purpose protein foundation models.
DPLM addresses tasks such as:
- Unconditional generation
- Motif scaffolding
- Structure-conditioned inverse folding
- Property-guided generation
- Downstream prediction
This work demonstrated a direction toward unifying sequence generation and protein representation learning.
From a historical perspective, the progression is natural:
EvoDiff → DPLM
EvoDiff can therefore be understood as a study that clearly established protein sequence diffusion as a meaningful research direction.
9. DPLM-2: Learning Sequences and Structures Together
An important next step after DPLM was DPLM-2, presented at ICLR 2025. [8]
DPLM-2 converts three-dimensional protein structures into discrete tokens and models both amino acid sequences and structural information within a single diffusion foundation model.
Conceptually, it aims to learn a joint distribution:
$$ P(\text{sequence},\text{structure}) $$
This makes it possible to address a variety of tasks within the same framework:
- Sequence → Structure
- Structure → Sequence
- Sequence + Structure → Completion
- Joint sequence–structure generation
- Motif scaffolding
This development is particularly representative of the broader research direction after EvoDiff.
EvoDiff demonstrated the potential of generating proteins directly from sequences.
DPLM expanded sequence diffusion into a general-purpose foundation model.
DPLM-2 then pursued greater versatility by jointly learning sequence and structure.
In other words, the sequence-first philosophy was extended while structural information was brought back into the model.
10. ESM3: Integrating Sequence, Structure, and Function
Another major model in this progression is ESM3.
The ESM3 research was published in Science in 2025. [9]
ESM3 models protein sequence, structure, and function within a unified generative framework.
$$ \text{Sequence}+\text{Structure}+\text{Function} $$
This represents a further step beyond sequence-only and structure-only models.
Generating a Fluorescent Protein with ESM3
One of the best-known achievements of ESM3 is the generation of a fluorescent protein called esmGFP.
Despite having a sequence identity of only approximately 58% to known fluorescent proteins, esmGFP was experimentally confirmed to exhibit fluorescence.
The researchers estimated the evolutionary distance represented by this sequence to be equivalent to approximately 500 million years. [9]
This result highlights the possibility that generative models can learn constraints from evolutionary data and explore sequence regions not observed in nature.
The Shared Philosophy of EvoDiff and ESM3
EvoDiff and ESM3 share an important underlying philosophy.
Both aim to learn from the enormous amount of information accumulated through natural evolution and use that knowledge to generate new proteins.
However, there is a clear difference between the two models.
EvoDiff focuses on protein sequences.
ESM3, by contrast, models sequences together with structures and functions.
This illustrates the relationship between EvoDiff's sequence-first philosophy and ESM3's multimodal approach.
11. 2025–2026: From Diffusion to Flow Matching and All-Atom Generation
During 2025–2026, protein generation research increasingly began modeling proteins at a finer level of detail.
The broad progression can be summarized as follows:
Around 2023: Backbone Generation
Around 2024–2025: Backbone + Sequence Generation
Around 2025–2026: All Atoms + Sequence + Target Interaction
In other words, researchers have moved toward generating not only protein backbones, but also amino acid sequences, side-chain atoms, and interactions with target molecules.
Proteina: Structure Generation with Flow Matching
One representative example of this direction is NVIDIA's Proteina.
Proteina is a protein backbone generation model based on flow matching, a generative modeling framework closely related to diffusion models.
The study expanded training data to include tens of millions of synthetic structures and reported the generation of designable backbones containing up to approximately 800 residues.
Proteina was selected as an Oral presentation at ICLR 2025. [10]
This research represents an effort to improve three-dimensional structure generation through large-scale data and alternative generative modeling techniques.
12. La-Proteina: Joint Generation of Sequences, Backbones, and Side Chains
In 2026, La-Proteina emerged as another important development.
La-Proteina is an all-atom protein generation model presented at ICLR 2026. [11]
It jointly generates:
- Amino acid sequences
- Protein backbones
- Side-chain atoms
The model represents Cα backbones using explicit three-dimensional coordinates while encoding amino acid identities and side-chain geometry into latent representations.
Generation is then performed using partially latent flow matching.
Toward Fully Atomistic Generation
La-Proteina reports fully atomistic joint sequence–structure generation for proteins containing up to 800 residues, along with active-site motif scaffolding.
Comparing this study with EvoDiff reveals a striking contrast.
EvoDiff adopts:
$$ \boxed{\text{Sequence Only}} $$
La-Proteina instead adopts:
$$ \boxed{\text{Sequence + Every Atom}} $$
This illustrates how protein generation research has expanded from representing proteins primarily as sequences to jointly generating sequences and structures at the atomic level.
13. A Major Research Direction in 2026: Binder Design
One of the particularly important topics in protein generation research in 2026 is binder design.
Binder design aims to create proteins that bind to specific target molecules.
Conceptually, the task can be represented as:
$$ \text{Target}\rightarrow\text{Protein Binder} $$
This is an important application area directly connected to drug discovery and the development of protein therapeutics.
Proteina-Complexa: Integrating Generation and Search
Proteina-Complexa, introduced in 2026, is a protein design approach based on fully atomistic generative modeling.
It combines generation with test-time search and optimization, using measures such as structure-prediction scores to evaluate generated candidates.
In simpler generative workflows, the primary operation is sampling candidates from a trained model.
Proteina-Complexa instead emphasizes a process that can be conceptually described as:
Generate → Evaluate → Search → Improve
The goal is not merely to produce candidate designs, but to improve them through evaluation and search.
The authors position their work as an integration of two approaches that have traditionally been considered distinct:
- Generative design
- Structure-predictor hallucination and search
Proteina-Complexa was selected as an Oral presentation at ICLR 2026. [12]
This development shows that protein generation models are evolving beyond systems that simply output new sequences or structures.
They are increasingly becoming design systems that combine generation, evaluation, and optimization.
14. BoltzGen: Integrating Structure Prediction and Protein Design
Another notable study as of 2026 is BoltzGen.
An updated version of the BoltzGen bioRxiv preprint was released in June 2026. [13]
BoltzGen uses an all-atom diffusion model to integrate structure prediction and protein design.
$$ \boxed{\text{Structure Prediction + Protein Design}} $$
Rather than focusing exclusively on protein–protein interactions, BoltzGen extends its design capabilities to a broader range of molecules and targets.
These include:
- Proteins
- Peptides
- Nanobodies
- Small molecules
- Disordered protein regions
The Importance of Wet-Lab Validation
One of the distinguishing features of BoltzGen is the extent of its experimental validation.
The study reports eight experimental campaigns spanning 26 targets.
It also demonstrates binders with nanomolar-level binding affinities against challenging targets. [13]
This reflects an important change in how protein generation research is evaluated.
Around 2023, one of the important evaluation methods was determining whether a generated protein appeared likely to adopt the intended structure according to prediction models such as AlphaFold.
More recent studies increasingly emphasize experimentally producing designed proteins and measuring their binding properties and biological functions.
The shift can be summarized as follows:
Does the generated protein appear likely to fold correctly according to a structure prediction model?
is evolving toward:
Can the protein actually be produced, and does it exhibit the intended binding properties or biological function?
This transition is particularly important for moving protein generative AI toward practical applications in drug discovery and pharmaceutical development.
15. EvoDiff Itself Was Also Updated in 2026
An often-overlooked point about EvoDiff is that the research did not end with the original 2023 preprint.
A new version of EvoDiff was published as an eLife Reviewed Preprint on August 24, 2026. [14]
This version includes experimental validation in addition to computational evaluations.
Experimental Evaluation
The evaluated aspects include:
- Expression of generated proteins
- Protein folding
- Secondary structure
- Intrinsically disordered mitochondrial targeting signals
- Metal-binding proteins
- Protein binders
The eLife assessment also characterizes EvoDiff as an order-agnostic diffusion foundation model for protein sequence generation. [14]
Thus, although EvoDiff was initially introduced in 2023, its experimental evaluation continued in subsequent years.
For readers studying EvoDiff in 2026, it is therefore important to examine not only the original 2023 preprint but also the 2026 Reviewed Preprint.
16. A Unified Historical View of Protein Generation Research
The studies discussed so far can be organized into a broad historical progression.
DeepSequence / UniRep
These studies demonstrated that biologically meaningful constraints and representations could be learned from evolutionary protein sequence data.
↓
ProGen / ProtGPT / ESM
Protein sequences began to be treated as language, enabling direct sequence generation and large-scale representation learning.
↓
ProteinMPNN
Inverse folding became highly practical, allowing amino acid sequences to be designed from three-dimensional backbone structures.
↓
RFdiffusion / Chroma (2023)
Diffusion models enabled the direct generation of three-dimensional protein backbones.
↓
EvoDiff (2023)
By applying diffusion to sequences instead of three-dimensional structures, EvoDiff demonstrated a way to directly leverage large-scale evolutionary sequence data.
↓
DPLM (2024)
Sequence diffusion was extended into a more general-purpose protein language model that supports both generation and representation learning.
↓
ESM3 / DPLM-2 (2024–2025)
Research progressed toward multimodal models that jointly generate sequences, structures, and functions.
↓
Proteina / La-Proteina (2025–2026)
Flow matching enabled advances in large-scale backbone generation and all-atom protein generation.
↓
BoltzGen / Proteina-Complexa (2026)
Research advanced toward all-atom design involving target interactions, with increasing emphasis on test-time search and wet-lab validation.
Comparing the Research Directions
| Period | Representative Models | Main Generation or Learning Targets | Research Direction |
|---|---|---|---|
| Around 2017–2019 | DeepSequence, UniRep | Protein sequences | Evolutionary representation learning and statistical modeling |
| Around 2020–2023 | ProGen, ProtGPT, ESM | Protein sequences | Direct generation and learning with protein language models |
| Around 2022–2023 | ProteinMPNN | Sequences conditioned on structures | Inverse folding |
| 2023 | RFdiffusion, Chroma | 3D backbones | Structure-based diffusion generation |
| 2023 | EvoDiff | Evolutionary sequences | Sequence-first diffusion |
| 2024 | DPLM | Sequences | Sequence diffusion foundation models |
| 2024–2025 | ESM3, DPLM-2 | Sequences, structures, and functions | Multimodal generation |
| 2025–2026 | Proteina, La-Proteina | Backbones, sequences, and atomic structures | Flow matching and all-atom generation |
| 2026 | BoltzGen, Proteina-Complexa | Atomic structures and target interactions | Target-conditioned design, integrated generation and search, and experimental validation |
This progression does not mean that each new technology simply replaced the previous one.
Rather, sequence-based and structure-based approaches evolved in parallel, and more recent research has increasingly sought to integrate their strengths.
17. Conclusion: Where Does EvoDiff Fit into the History of Protein Generation?
The historical position of EvoDiff can be summarized as follows:
EvoDiff emerged in 2023, when protein diffusion was becoming closely associated with three-dimensional structure generation, and established sequence-first protein design as an alternative direction by applying diffusion directly to evolutionary protein sequences.
The ideas introduced by EvoDiff were subsequently developed in models such as DPLM, which expanded sequence diffusion into a more general-purpose foundation modeling framework.
Meanwhile, the current frontier has moved beyond sequence-only generation toward multimodal and all-atom generative models that integrate protein sequences, structures, functions, and molecular interactions.
Therefore, although EvoDiff is historically important, it would be inaccurate to conclude that sequence-only modeling represents the most effective protein design strategy as of 2026.
Instead, the central development is the gradual convergence of two important ideas:
- EvoDiff's approach: Leveraging the enormous scale of evolutionary sequence data.
- RFdiffusion's approach: Directly modeling three-dimensional molecular geometry.
Models such as DPLM-2, ESM3, La-Proteina, and BoltzGen illustrate how these ideas are increasingly being integrated.
In addition, the evaluation of protein generative AI has evolved beyond computational structure prediction toward experimental verification of protein expression, binding affinity, and biological function.
This transition is especially relevant to drug discovery and the development of protein therapeutics.
Ultimately, EvoDiff's historical contribution was not to prove that "sequence is all you need."
Its contribution was to demonstrate that protein sequences themselves can serve as the generative space, opening a research direction distinct from structure generation and helping establish a foundation for subsequent integrated protein generative models.
References
[1] Riesselman et al., "Deep generative models of genetic variation capture the effects of mutations," Nature Methods, 15, 816–822, 2018. https://www.nature.com/articles/s41592-018-0138-4.pdf
[2] Alley et al., "Unified rational protein engineering with sequence-based deep representation learning," Nature Methods, 16, 1315–1322, 2019. https://www.nature.com/articles/s41592-019-0598-1
[3] Madani et al., "Large language models generate functional protein sequences across diverse families," Nature Biotechnology, 41, 1099–1106, 2023. https://preview-www.nature.com/articles/s41587-022-01618-2.pdf
[4] Ingraham et al., "Illuminating protein space with a programmable generative model," Nature, 623, 1070–1078, 2023. https://www.nature.com/articles/s41586-023-06728-8
[5] Alamdari et al., "Protein generation with evolutionary diffusion: sequence is all you need," bioRxiv, 2023. https://www.biorxiv.org/content/10.1101/2023.09.11.556673v1
[6] Alamdari et al., "Protein generation with evolutionary diffusion: sequence is all you need," bioRxiv, revised preprint, 2023. https://www.biorxiv.org/content/10.1101/2023.09.11.556673v2.full
[7] Wang et al., "DPLM: A Diffusion Protein Language Model," Proceedings of the 41st International Conference on Machine Learning (ICML), 2024. https://arxiv.org/abs/2402.18567
[8] "DPLM-2: A Multimodal Diffusion Protein Language Model," International Conference on Learning Representations (ICLR), 2025. https://proceedings.iclr.cc/paper_files/paper/2025/hash/57c30b677add9aa78e1745f0643104d0-Abstract-Conference.html
[9] Hayes et al., "Simulating 500 million years of evolution with a language model," Science, 387, 850–858, 2025. https://pubmed.ncbi.nlm.nih.gov/39818825/
[10] "Proteina: Scaling Flow-based Protein Structure Generative Models," International Conference on Learning Representations (ICLR), 2025. https://proceedings.iclr.cc/paper_files/paper/2025/hash/f4e9121ad30cd4e5528042fbfd835b3f-Abstract-Conference.html
[11] "La-Proteina: Atomistic Protein Generation with Partially Latent Flow Matching," International Conference on Learning Representations (ICLR), 2026. https://research.nvidia.com/publication/2026-01_la-proteina-atomistic-protein-generation-partially-latent-flow-matching
[12] "Proteina-Complexa," International Conference on Learning Representations (ICLR), 2026. https://proceedings.iclr.cc/paper_files/paper/2026/hash/639d992f819c2b40387d4d5170b8ffd7-Abstract-Conference.html
[13] "BoltzGen," bioRxiv, revised preprint, 2026. https://www.biorxiv.org/content/10.1101/2025.11.20.689494v2.full
[14] Alamdari et al., "Protein generation with evolutionary diffusion: sequence is all you need," eLife, Reviewed Preprint, 2026. https://elifesciences.org/reviewed-preprints/112029.pdf