Dataset Viewer
id
int64 1
361k
| content
stringlengths 1
23.3M
⌀ | title
stringlengths 1
271
| appendix
bool 1
class | paper_arxiv_id
stringlengths 10
13
| section_in_paper_id
float64 1
294
⌀ |
|---|---|---|---|---|---|
95,360
|
\label{sec:conclusion}
In this work, we present DiffSpectra, a novel spectrum-conditioned diffusion framework for molecular structure elucidation. By integrating a continuous-time variational diffusion model with a Transformer-based molecular graph backbone, DiffSpectra is capable of generating molecular graphs with consistent 2D and 3D structures under spectral conditions. To encode spectral conditions, we introduce SpecFormer, a multi-spectrum Transformer encoder, which is further pre-trained via masked patch reconstruction and contrastive learning to align spectra with molecular structures. Through extensive experiments, we demonstrate that DiffSpectra achieves superior performance in recovering molecular structures from spectra, benefiting from the incorporation of pre-trained SpecFormer and multi-modal spectral inputs. In addition, our analysis shows that a model-based SE(3) equivariant architecture provides robust geometric consistency, while proper sampling temperature helps balance stochasticity and accuracy in generation.
Our work paves the way for several promising directions for future research. Firstly, expanding the diversity and quantity of molecular spectra data could further boost the performance of spectral-conditioned generation. Secondly, extending the framework to handle even more complex spectra modalities, such as NMR or mass spectra, could generalize the method toward broader applications in analytical chemistry. Lastly, adapting DiffSpectra to larger biomolecular or materials systems may open opportunities for molecular identification in drug discovery and materials science, especially when combined with high-throughput experimental pipelines. We believe DiffSpectra represents a significant step toward trustworthy and accurate molecular structure elucidation conditioned on spectral data.
|
Conclusion
| false
|
2507.06853
| 3
|
95,361
|
\label{sec:method}
\subsection{Diffusion Model for Molecular Structure Elucidation}
We adopt a continuous-time diffusion probabilistic model to generate molecular graphs in a joint space that integrates 2D topology and 3D geometry. Each molecule is represented as a graph $\mathcal{G} = (\mathbf{H}, \mathbf{A}, \mathbf{X})$\mathcal{G} = (\mathbf{H}, \mathbf{A}, \mathbf{X}), where $\mathbf{H} \in \mathbb{R}^{N \times d_1}$\mathbf{H} \in \mathbb{R}^{N \times d_1}N \times d_1 denotes node-level attributes such as atom types and charges, $\mathbf{A} \in \mathbb{R}^{N \times N \times d_2}$\mathbf{A} \in \mathbb{R}^{N \times N \times d_2}N \times N \times d_2 encodes pairwise edge features such as bond existence and bond types, and $\mathbf{X} \in \mathbb{R}^{N \times 3}$\mathbf{X} \in \mathbb{R}^{N \times 3}N \times 3 corresponds to the 3D coordinates of the atoms. Here, $N$N represents the number of atoms in the molecule, $d_1$d_1 denotes the dimensionality of the node features, and $d_2$d_2 denotes the dimensionality of the edge features.
We adopt the mathematical formulation of the Variational Diffusion Model (VDM)~\citep{VDM,ProgressiveDistillation,BlurringDiffusion} to define our diffusion and sampling processes, which follow a continuous-time diffusion probabilistic framework, as illustrated in \cref{fig:overview}(A). To streamline the presentation, we consider a generic graph component, namely the node features $\mathbf{H}$\mathbf{H}, the edge features $\mathbf{A}$\mathbf{A}, or the atomic coordinates $\mathbf{X}$\mathbf{X}, which we uniformly denote by a vector-valued variable $\mathbf{G} \in \mathbb{R}^d$\mathbf{G} \in \mathbb{R}^d. The subsequent formulas can then be applied independently to each component in a consistent manner.
\subsubsection{Forward Diffusion Process}
In the forward diffusion process, Gaussian noise $\bG_\epsilon \sim \mathcal{N}(\b0, \bI)$\bG_\epsilon \sim \mathcal{N}(\b0, \bI) is gradually added to the data over continuous time from $t=0$t=0 to $t=1$t=1. The noised sample $\bG_t$\bG_t is obtained by:
\begin{equation}
\bG_t = \alpha_t \bG_0 + \sigma_t \bG_\epsilon,
\end{equation}\begin{equation}
\bG_t = \alpha_t \bG_0 + \sigma_t \bG_\epsilon,
\end{equation}
\bG_t = \alpha_t \bG_0 + \sigma_t \bG_\epsilon,
where $\alpha_t$\alpha_t and $\sigma_t$\sigma_t are signal and noise scaling functions, typically determined by cosine or linear schedules. The signal-to-noise ratio (SNR) at time $t$t is defined as:
\begin{equation}
\mathrm{SNR}(t) = \frac{\alpha_t^2}{\sigma_t^2}.
\end{equation}\begin{equation}
\mathrm{SNR}(t) = \frac{\alpha_t^2}{\sigma_t^2}.
\end{equation}
\mathrm{SNR}(t) = \frac{\alpha_t^2}{\sigma_t^2}.
The SNR is strictly decreasing from $t = 0$t = 0 to $t = 1$t = 1, ensuring that the data is gradually corrupted into pure noise as $t \to 1$t \to 1.
Given two times $0 \leq s < t \leq 1$0 \leq s < t \leq 1, the conditional distribution of a noised graph at time $t$t given a less-noised graph at time $s$s, denoted $q(\bG_t \mid \bG_s)$q(\bG_t \mid \bG_s), is also Gaussian:
\begin{equation}
q(\bG_t \mid \bG_s) = \mathcal{N}(\alpha_{t|s} \bG_s, \sigma_{t|s}^2 \bI),
\end{equation}\begin{equation}
q(\bG_t \mid \bG_s) = \mathcal{N}(\alpha_{t|s} \bG_s, \sigma_{t|s}^2 \bI),
\end{equation}
q(\bG_t \mid \bG_s) = \mathcal{N}(\alpha_{t|s}t|s \bG_s, \sigma_{t|s}t|s^2 \bI),
where the intermediate scaling factor and conditional variance are defined as:
\begin{equation}
\alpha_{t|s} = \frac{\alpha_t}{\alpha_s}, \quad
\sigma_{t|s}^2 = \sigma_t^2 - \alpha_{t|s}^2 \sigma_s^2.
\end{equation}\begin{equation}
\alpha_{t|s} = \frac{\alpha_t}{\alpha_s}, \quad
\sigma_{t|s}^2 = \sigma_t^2 - \alpha_{t|s}^2 \sigma_s^2.
\end{equation}
\alpha_{t|s}t|s = \frac{\alpha_t}{\alpha_s}, \quad
\sigma_{t|s}t|s^2 = \sigma_t^2 - \alpha_{t|s}t|s^2 \sigma_s^2.
This formulation allows the forward process to be implemented using analytically tractable Gaussian transitions, which also facilitates efficient computation of training objectives and reverse-time sampling.
\subsubsection{Reverse Denoising Process}
The reverse-time denoising process starts from a standard Gaussian sample $\bG_1 \sim \mathcal{N}(\b0, \bI)$\bG_1 \sim \mathcal{N}(\b0, \bI) and proceeds backward from $t = 1$t = 1 to $t = 0$t = 0. Let $0 \leq s < t \leq 1$0 \leq s < t \leq 1 denote two consecutive time steps. At each step, a data prediction model $d_\theta$d_\theta takes as input the noised sample $\bG_t$\bG_t, a self-conditioning~\cite{SelfConditioning} estimate $\tilde{\bG}_0$\tilde{\bG}_0, and the log signal-to-noise ratio $\log \mathrm{SNR}(t)$\log \mathrm{SNR}(t). The model predicts $\hat{\bG}_0 = d_\theta(\bG_t, \tilde{\bG}_0, \log \mathrm{SNR}(t))$\hat{\bG}_0 = d_\theta(\bG_t, \tilde{\bG}_0, \log \mathrm{SNR}(t)), which is then used to compute the denoised sample $\bG_s$\bG_s by:
\begin{gather}
\bar{\bG}_s = \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \hat{\bG}_0, \\
\bG_s = \bar{\bG}_s + \tau \cdot \frac{\sigma_s \sigma_{t|s}}{\sigma_t} \bG_\epsilon, \quad \bG_\epsilon \sim \mathcal{N}(\b0, \bI),
\label{eq:sampling}
\end{gather}\begin{gather}
\bar{\bG}_s = \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \hat{\bG}_0, \\
\bG_s = \bar{\bG}_s + \tau \cdot \frac{\sigma_s \sigma_{t|s}}{\sigma_t} \bG_\epsilon, \quad \bG_\epsilon \sim \mathcal{N}(\b0, \bI),
\label{eq:sampling}
\end{gather}
\bar{\bG}_s = \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \hat{\bG}_0, \\
\bG_s = \bar{\bG}_s + \tau \cdot \frac{\sigma_s \sigma_{t|s}}{\sigma_t} \bG_\epsilon, \quad \bG_\epsilon \sim \mathcal{N}(\b0, \bI),
\label{eq:sampling}
where $\tau$\tau is a sampling temperature parameter, modulating the amount of stochasticity in the sampling process. Lower values of $\tau < 1$\tau < 1 reduce the influence of noise, leading to more deterministic outputs.
The self-conditioning mechanism, where the model reuses its previous prediction $\tilde{\bG}_0$\tilde{\bG}_0 as an additional input in the next step, enhances training stability and generation quality~\cite{SelfConditioning}.
\subsubsection{Training Objective}
To recover the original sample, we train a data prediction model $d_\theta$d_\theta that takes as input the noised graph $\mathbf{G}_t$\mathbf{G}_t, a self-conditioning estimate $\tilde{\mathbf{G}}_0$\tilde{\mathbf{G}}_0, and the log signal-to-noise ratio $\log \mathrm{SNR}(t)$\log \mathrm{SNR}(t). The model predicts the clean sample $\hat{\mathbf{G}}_0$\hat{\mathbf{G}}_0, where $\hat{\mathbf{G}}_0$\hat{\mathbf{G}}_0 refers to any component of $\hat{\mathcal{G}}_0 = (\hat{\mathbf{H}}, \hat{\mathbf{A}}, \hat{\mathbf{X}})$\hat{\mathcal{G}}_0 = (\hat{\mathbf{H}}, \hat{\mathbf{A}}, \hat{\mathbf{X}}). Therefore, the model is trained to minimize a weighted mean squared error:
\begin{equation}
\mathcal{L} = \mathbb{E}_{t, \mathbf{G}_0} \left[ \sqrt{\frac{\alpha_t}{\sigma_t}} \left( \lambda_1 \| \hat{\mathbf{A}} - \mathbf{A}_0 \|_2^2 + \lambda_2 \| \hat{\mathbf{X}} - \hat{\mathbf{X}}_0 \|_2^2 + \lambda_3 \| \hat{\mathbf{H}} - \mathbf{H}_0 \|_2^2 \right) \right],
\end{equation}\begin{equation}
\mathcal{L} = \mathbb{E}_{t, \mathbf{G}_0} \left[ \sqrt{\frac{\alpha_t}{\sigma_t}} \left( \lambda_1 \| \hat{\mathbf{A}} - \mathbf{A}_0 \|_2^2 + \lambda_2 \| \hat{\mathbf{X}} - \hat{\mathbf{X}}_0 \|_2^2 + \lambda_3 \| \hat{\mathbf{H}} - \mathbf{H}_0 \|_2^2 \right) \right],
\end{equation}
\mathcal{L} = \mathbb{E}_{t, \mathbf{G}_0}t, \mathbf{G}_0 \left[ \sqrt{\frac{\alpha_t}{\sigma_t}} \left( \lambda_1 \| \hat{\mathbf{A}} - \mathbf{A}_0 \|_2^2 + \lambda_2 \| \hat{\mathbf{X}} - \hat{\mathbf{X}}_0 \|_2^2 + \lambda_3 \| \hat{\mathbf{H}} - \mathbf{H}_0 \|_2^2 \right) \right],
where $\hat{\mathbf{X}}_0$\hat{\mathbf{X}}_0 is obtained by aligning $\mathbf{X}_0$\mathbf{X}_0 to $\mathbf{X}_t$\mathbf{X}_t using the Kabsch algorithm~\cite{KabschAlign} to preserve SE(3)-equivariance. The loss coefficients $\lambda_i$\lambda_i balance the relative importance of each component.
\subsection{Diffusion Molecule Transformer (DMT)}
To parameterize the data prediction module $d_\theta$d_\theta, we employ a specialized architecture named \emph{Diffusion Molecule Transformer} (DMT)~\cite{JODO,NExT-Mol} as its backbone, to jointly model the noisy molecular graph’s node features, edge features, and 3D coordinates in our diffusion framework. The architecture is motivated by the need to recover the correlations among these three components, which are independently corrupted by noise during the forward diffusion process.
DMT is specifically tailored for molecular generation tasks and respects both permutation equivariance and SE(3)-equivariance.
At each time $t \in [0,1]$t \in [0,1] during the reverse process, DMT takes the noisy molecular graph $\mathcal{G}_t = (\mathbf{H}_t, \mathbf{A}_t, \mathbf{X}_t)$\mathcal{G}_t = (\mathbf{H}_t, \mathbf{A}_t, \mathbf{X}_t) as input, along with a self-conditioned estimate $\tilde{\mathcal{G}}_0$\tilde{\mathcal{G}}_0, and predicts a clean graph $\hat{\mathcal{G}}_0 = (\hat{\mathbf{H}}, \hat{\mathbf{A}}, \hat{\mathbf{X}})$\hat{\mathcal{G}}_0 = (\hat{\mathbf{H}}, \hat{\mathbf{A}}, \hat{\mathbf{X}}). The model performs denoising inference conditioned on both the current time step and the molecular spectra. The overall model architecture is illustrated in \cref{fig:overview}(B). For the sake of clarity, the self-conditioning mechanism is omitted in the figure.
A time embedding, obtained via sinusoidal encoding of $\log(\alpha_t^2/\sigma_t^2)$\log(\alpha_t^2/\sigma_t^2), is passed through learnable projections. Simultaneously, a spectra embedding is generated by the SpecFormer, which will be detailed in \cref{sec:specformer}. These two embeddings are then concatenated to form the conditioning vector $\bC$\bC, which is used throughout the network.
\textbf{Initial embeddings.}
The DMT module is composed of stacked Transformer-style equivariant blocks. In the first layer, the node features $\bH_t$\bH_t and the edge features $\bA_t$\bA_t are first concatenated with their corresponding self-conditioning estimates $\tilde{\bH}_0$\tilde{\bH}_0 and $\tilde{\bA}_0$\tilde{\bA}_0, respectively. For edge features, distance features $\mathbf{D}_0$\mathbf{D}_0 computed from the self-conditioned coordinates $\tilde{\bX}_0$\tilde{\bX}_0 are also concatenated. These concatenated features are then passed through linear projections to obtain the initial embeddings:
\begin{equation}
\bH^{(1)} = \mathrm{Linear}_{\mathrm{node}}\bigl(\left[\bH_t, \tilde{\bH}_0\right]\bigr) \in \mathbb{R}^{N \times d_{h}},
\end{equation}\begin{equation}
\bH^{(1)} = \mathrm{Linear}_{\mathrm{node}}\bigl(\left[\bH_t, \tilde{\bH}_0\right]\bigr) \in \mathbb{R}^{N \times d_{h}},
\end{equation}
\bH^{(1)}(1) = \mathrm{Linear}_{\mathrm{node}}\mathrm{node}\bigl(\left[\bH_t, \tilde{\bH}_0\right]\bigr) \in \mathbb{R}^{N \times d_{h}}N \times d_{h}h,
\begin{equation}
\bA^{(1)} = \mathrm{Linear}_{\mathrm{edge}}\bigl(\left[\bA_t, \tilde{\bA}_0, \mathbf{D}_0\right]\bigr) \in \mathbb{R}^{N \times N \times d_{a}},
\end{equation}\begin{equation}
\bA^{(1)} = \mathrm{Linear}_{\mathrm{edge}}\bigl(\left[\bA_t, \tilde{\bA}_0, \mathbf{D}_0\right]\bigr) \in \mathbb{R}^{N \times N \times d_{a}},
\end{equation}
\bA^{(1)}(1) = \mathrm{Linear}_{\mathrm{edge}}\mathrm{edge}\bigl(\left[\bA_t, \tilde{\bA}_0, \mathbf{D}_0\right]\bigr) \in \mathbb{R}^{N \times N \times d_{a}}N \times N \times d_{a}a,
where $\left[\cdot,\cdot\right]$\left[\cdot,\cdot\right] denotes the concatenation operation, and $d_{h}$d_{h}h and $d_{a}$d_{a}a denote the hidden embedding dimensions for node and edge embeddings, respectively.
The coordinate stream initializes with the noisy positions $\bX^{(1)} = \bX_t$\bX^{(1)}(1) = \bX_t. In addition, adjacency matrices derived from chemical bonding patterns and distance cutoffs computed from $\tilde{\bX}_0$\tilde{\bX}_0 are extracted and concatenated to serve as auxiliary attention masks in subsequent layers.
\textbf{Three-stream update.}
Each DMT block consists of three interacting streams that are responsible for updating node features, edge features, and coordinates, respectively. These streams exchange information throughout the block. In the node stream, information is propagated across the molecular graph via a relational multi-head attention (MHA) mechanism~\cite{GTSurvey,GSLB}, enabling flexible message passing over fully connected molecular graphs. In particular, pairwise geometric distances computed from coordinates are transformed into radial basis encodings $\rho_{ij}^{(l)}$\rho_{ij}ij^{(l)}(l), and these are combined with edge features and geometric distances such as:
\begin{equation}
\bar{\bA}_{ij}^{(l)} = \left[ \bA_{ij}^{(l)};\, \|\bX^{(l)}_i - \bX^{(l)}_j\|_2;\, \rho_{ij}^{(l)} \right],
\end{equation}\begin{equation}
\bar{\bA}_{ij}^{(l)} = \left[ \bA_{ij}^{(l)};\, \|\bX^{(l)}_i - \bX^{(l)}_j\|_2;\, \rho_{ij}^{(l)} \right],
\end{equation}
\bar{\bA}_{ij}ij^{(l)}(l) = \left[ \bA_{ij}ij^{(l)}(l);\, \|\bX^{(l)}(l)_i - \bX^{(l)}(l)_j\|_2;\, \rho_{ij}ij^{(l)}(l) \right],
forming a geometry-aware relational representation. The queries, keys, and values for MHA are obtained by applying learnable linear projections to node features:
\begin{equation}
\bQ = \bH^{(l)} \bW^Q, \quad
\bK = \bH^{(l)} \bW^K, \quad
\bV = \bH^{(l)} \bW^V,
\end{equation}\begin{equation}
\bQ = \bH^{(l)} \bW^Q, \quad
\bK = \bH^{(l)} \bW^K, \quad
\bV = \bH^{(l)} \bW^V,
\end{equation}
\bQ = \bH^{(l)}(l) \bW^Q, \quad
\bK = \bH^{(l)}(l) \bW^K, \quad
\bV = \bH^{(l)}(l) \bW^V,
where $\bW^Q$\bW^Q, $\bW^K$\bW^K, and $\bW^V$\bW^V are learned projection matrices. The attention weights are then defined as:
\begin{equation}
a_{ij} = \frac{ \tanh\bigl(\phi_0(\bar{\bA}_{ij}^{(l)})\bigr)\, \bQ_i\, \bK_j^\top }{ \sqrt{d_k} },
\quad
a = \mathrm{softmax}(a),
\end{equation}\begin{equation}
a_{ij} = \frac{ \tanh\bigl(\phi_0(\bar{\bA}_{ij}^{(l)})\bigr)\, \bQ_i\, \bK_j^\top }{ \sqrt{d_k} },
\quad
a = \mathrm{softmax}(a),
\end{equation}
a_{ij}ij = \frac{ \tanh\bigl(\phi_0(\bar{\bA}_{ij}^{(l)})\bigr)\, \bQ_i\, \bK_j^\top }{ \sqrt{d_k} },
\quad
a = \mathrm{softmax}(a),
and the node-level aggregation proceeds as:
\begin{equation}
\mathrm{Attn}(\bH^{(l)}, \bar{\bA}^{(l)})_i
= \sum_{j=1}^N a_{ij}\, \tanh\bigl(\phi_1(\bar{\bA}_{ij}^{(l)})\bigr)\, \bV_j,
\end{equation}\begin{equation}
\mathrm{Attn}(\bH^{(l)}, \bar{\bA}^{(l)})_i
= \sum_{j=1}^N a_{ij}\, \tanh\bigl(\phi_1(\bar{\bA}_{ij}^{(l)})\bigr)\, \bV_j,
\end{equation}
\mathrm{Attn}(\bH^{(l)}(l), \bar{\bA}^{(l)}(l))_i
= \sum_{j=1}j=1^N a_{ij}ij\, \tanh\bigl(\phi_1(\bar{\bA}_{ij}ij^{(l)}(l))\bigr)\, \bV_j,
with $\phi_0$\phi_0 and $\phi_1$\phi_1 as learnable projections.
We extend the above attention mechanism to a multi-head framework, yielding the aggregated representation $\bM^{(l)} \in \mathbb{R}^{N \times d_m}$\bM^{(l)}(l) \in \mathbb{R}^{N \times d_m}N \times d_m, which serves as the intermediate node embeddings after relational multi-head attention.
To incorporate time-dependent and spectra-dependent conditioning, we apply adaptive layer normalization (AdaLN) to each stream, conditioned on a learned conditioning embedding $\mathbf{C}$\mathbf{C}: $\mathrm{AdaLN}(\vh, \mathbf{C}) = \left( 1 + \mathrm{FFN}_{\mathrm{scale}}(\mathbf{C}) \right) \cdot \mathrm{LN}(\vh) + \mathrm{FFN}_{\mathrm{bias}}(\mathbf{C})$\mathrm{AdaLN}(\vh, \mathbf{C}) = \left( 1 + \mathrm{FFN}_{\mathrm{scale}}\mathrm{scale}(\mathbf{C}) \right) \cdot \mathrm{LN}(\vh) + \mathrm{FFN}_{\mathrm{bias}}\mathrm{bias}(\mathbf{C}),
and use an adaptive scaling function: $\mathrm{Scale}(\vh, \mathbf{C}) = \mathrm{FFN}_{\mathrm{scale}}'(\mathbf{C}) \cdot \vh$\mathrm{Scale}(\vh, \mathbf{C}) = \mathrm{FFN}_{\mathrm{scale}}\mathrm{scale}'(\mathbf{C}) \cdot \vh
to modulate feature amplitudes. The node update is performed by:
\begin{equation}
\bH^{(l+1)'} = \mathrm{Scale}\bigl(\bM^{(l)}, \mathbf{C}\bigr) + \bH^{(l)},
\end{equation}\begin{equation}
\bH^{(l+1)'} = \mathrm{Scale}\bigl(\bM^{(l)}, \mathbf{C}\bigr) + \bH^{(l)},
\end{equation}
\bH^{(l+1)'}(l+1)' = \mathrm{Scale}\bigl(\bM^{(l)}(l), \mathbf{C}\bigr) + \bH^{(l)}(l),
\begin{equation}
\bH^{(l+1)} = \mathrm{Scale}\left(
\mathrm{FFN}\bigl( \mathrm{AdaLN}(\bH^{(l+1)'}, \mathbf{C})\bigr),
\mathbf{C}
\right) + \bH^{(l+1)'}.
\end{equation}\begin{equation}
\bH^{(l+1)} = \mathrm{Scale}\left(
\mathrm{FFN}\bigl( \mathrm{AdaLN}(\bH^{(l+1)'}, \mathbf{C})\bigr),
\mathbf{C}
\right) + \bH^{(l+1)'}.
\end{equation}
\bH^{(l+1)}(l+1) = \mathrm{Scale}\left(
\mathrm{FFN}\bigl( \mathrm{AdaLN}(\bH^{(l+1)'}(l+1)', \mathbf{C})\bigr),
\mathbf{C}
\right) + \bH^{(l+1)'}(l+1)'.
Similarly, edge features are updated by first fusing node messages:
\begin{equation}
\hat{\bA}_{ij}^{(l)} = \left( \bM^{(l)}_i + \bM^{(l)}_j \right) \bW_1,
\end{equation}\begin{equation}
\hat{\bA}_{ij}^{(l)} = \left( \bM^{(l)}_i + \bM^{(l)}_j \right) \bW_1,
\end{equation}
\hat{\bA}_{ij}ij^{(l)}(l) = \left( \bM^{(l)}(l)_i + \bM^{(l)}(l)_j \right) \bW_1,
with subsequent normalization and scaling:
\begin{equation}
\bA^{(l+1)'}_{ij} = \mathrm{Scale}\bigl( \hat{\bA}_{ij}^{(l)}, \mathbf{C} \bigr) + \bA^{(l)}_{ij},
\end{equation}\begin{equation}
\bA^{(l+1)'}_{ij} = \mathrm{Scale}\bigl( \hat{\bA}_{ij}^{(l)}, \mathbf{C} \bigr) + \bA^{(l)}_{ij},
\end{equation}
\bA^{(l+1)'}(l+1)'_{ij}ij = \mathrm{Scale}\bigl( \hat{\bA}_{ij}ij^{(l)}(l), \mathbf{C} \bigr) + \bA^{(l)}(l)_{ij}ij,
\begin{equation}
\bA^{(l+1)}_{ij} = \mathrm{Scale}\left(
\mathrm{FFN}\bigl( \mathrm{AdaLN}(\bA^{(l+1)'}_{ij}, \mathbf{C})\bigr),
\mathbf{C}
\right) + \bA^{(l+1)'}_{ij}.
\end{equation}\begin{equation}
\bA^{(l+1)}_{ij} = \mathrm{Scale}\left(
\mathrm{FFN}\bigl( \mathrm{AdaLN}(\bA^{(l+1)'}_{ij}, \mathbf{C})\bigr),
\mathbf{C}
\right) + \bA^{(l+1)'}_{ij}.
\end{equation}
\bA^{(l+1)}(l+1)_{ij}ij = \mathrm{Scale}\left(
\mathrm{FFN}\bigl( \mathrm{AdaLN}(\bA^{(l+1)'}(l+1)'_{ij}ij, \mathbf{C})\bigr),
\mathbf{C}
\right) + \bA^{(l+1)'}(l+1)'_{ij}ij.
The parameter $\bW_1$\bW_1 is a learnable weight matrix that couples node and edge information.
For the coordinate stream, equivariant updates are achieved using directional vector fields, combining signals from node and edge streams:
\begin{equation}
e_{ij}^{(l+1)} = \mathrm{AdaLN}\left(
\bW_2
\left[
\bH_i^{(l+1)},
\bH_j^{(l+1)},
\bA_{ij}^{(l+1)},
\|\bX^{(l)}_i - \bX^{(l)}_j\|_2
\right],
\mathbf{C}
\right),
\end{equation}\begin{equation}
e_{ij}^{(l+1)} = \mathrm{AdaLN}\left(
\bW_2
\left[
\bH_i^{(l+1)},
\bH_j^{(l+1)},
\bA_{ij}^{(l+1)},
\|\bX^{(l)}_i - \bX^{(l)}_j\|_2
\right],
\mathbf{C}
\right),
\end{equation}
e_{ij}ij^{(l+1)}(l+1) = \mathrm{AdaLN}\left(
\bW_2
\left[
\bH_i^{(l+1)}(l+1),
\bH_j^{(l+1)}(l+1),
\bA_{ij}ij^{(l+1)}(l+1),
\|\bX^{(l)}(l)_i - \bX^{(l)}(l)_j\|_2
\right],
\mathbf{C}
\right),
\begin{equation}
\bX_i^{(l+1)} = \bX^{(l)}_i + \sum_{j \ne i} \gamma^{(l)} \,
\frac{ \bX^{(l)}_i - \bX^{(l)}_j }
{ \|\bX^{(l)}_i - \bX^{(l)}_j\|_2 }
\, \tanh\bigl(\mathrm{FFN}( e_{ij}^{(l+1)} )\bigr),
\end{equation}\begin{equation}
\bX_i^{(l+1)} = \bX^{(l)}_i + \sum_{j \ne i} \gamma^{(l)} \,
\frac{ \bX^{(l)}_i - \bX^{(l)}_j }
{ \|\bX^{(l)}_i - \bX^{(l)}_j\|_2 }
\, \tanh\bigl(\mathrm{FFN}( e_{ij}^{(l+1)} )\bigr),
\end{equation}
\bX_i^{(l+1)}(l+1) = \bX^{(l)}(l)_i + \sum_{j \ne i}j \ne i \gamma^{(l)}(l) \,
\frac{ \bX^{(l)}_i - \bX^{(l)}_j }
{ \|\bX^{(l)}_i - \bX^{(l)}_j\|_2 }
\, \tanh\bigl(\mathrm{FFN}( e_{ij}ij^{(l+1)}(l+1) )\bigr),
where $\bW_2$\bW_2 is a learned projection and $\gamma^{(l)}$\gamma^{(l)}(l) is a trainable scalar that stabilizes the directional updates. Finally, coordinates are shifted to have zero center-of-mass by mean centering, ensuring translation invariance.
Overall, the DMT stacks $L$L such equivariant blocks, which collectively refine the graph features and spatial coordinates. Final output heads predict discrete atom and bond types as well as coordinates aligned to the canonical centered frame. Its design guarantees permutation equivariance and SE(3) equivariance, enabling robust molecule generation across both topological and geometric levels.
\subsection{Spectra Transformer (SpecFormer) for Spectra Encoding}\label{sec:specformer}
To enable DMT to incorporate molecular spectral information as a conditional input for molecular structure elucidation, we propose the Spectra Transformer (SpecFormer)~\cite{MolSpectra} to encode molecular spectra. We further pre-train SpecFormer using both molecular structure data and spectra data.
\subsubsection{Architecture of SpecFormer}
For each type of spectrum, we first segment the data independently into patches and encode these patches. The patch embeddings from all spectra are then concatenated and collectively processed by a Transformer-based encoder.
\textbf{Patching.}
Rather than encoding each frequency point individually, we divide each spectrum into several patches.
This design choice offers two key benefits: (\romannumeral 1) It allows the model to capture local semantic features—such as absorption peaks—more effectively by grouping adjacent frequency points; and (\romannumeral 2) it lowers the computational cost for the following Transformer layers.
Formally, a spectrum $\vs_i \in \mathbb{R}^{L_i}$\vs_i \in \mathbb{R}^{L_i}L_i (where $i = 1, \dots, |\mathcal{S}|$i = 1, \dots, |\mathcal{S}|) is split into patches of length $P_i$P_i with stride $D_i$D_i. If $0 < D_i < P_i$0 < D_i < P_i, patches overlap by $P_i - D_i$P_i - D_i points; if $D_i = P_i$D_i = P_i, patches are non-overlapping. The patching process yields a sequence $\vp_i \in \mathbb{R}^{N_i \times P_i}$\vp_i \in \mathbb{R}^{N_i \times P_i}N_i \times P_i, where $N_i = \left\lfloor \frac{L_i - P_i}{D_i} \right\rfloor + 1$N_i = \left\lfloor \frac{L_i - P_i}{D_i} \right\rfloor + 1 is the number of patches produced from $\vs_i$\vs_i.
\textbf{Patch and position encoding.}
Each patch sequence from the $i$i-th spectrum is projected into a latent space of dimension $d$d via a learnable linear transformation $\bW_i \in \mathbb{R}^{P_i \times d}$\bW_i \in \mathbb{R}^{P_i \times d}P_i \times d. To preserve the sequential order, a learnable position encoding $\bW_i^{\text{pos}} \in \mathbb{R}^{N_i \times d}$\bW_i^{\text{pos}}\text{pos} \in \mathbb{R}^{N_i \times d}N_i \times d is added: $\vp_i' = \vp_i \bW_i + \bW_i^{\text{pos}}$\vp_i' = \vp_i \bW_i + \bW_i^{\text{pos}}\text{pos}, yielding the encoded representation $\vp_i' \in \mathbb{R}^{N_i \times d}$\vp_i' \in \mathbb{R}^{N_i \times d}N_i \times d for each spectrum. These representations serve as input to the SpecFormer encoder.
\textbf{SpecFormer: multi-spectrum Transformer encoder.}
Existing encoders such as CNN-AM~\citep{CNN-AM} utilize one-dimensional convolutions and are typically designed to process a single type of spectrum. In contrast, our model simultaneously considers multiple molecular spectra (e.g., UV-Vis, IR, Raman) as input. This multi-spectrum strategy is motivated by the presence of both \textit{intra-spectrum dependencies}—relationships among peaks within the same spectrum—and \textit{inter-spectrum dependencies}—correlations between peaks across different spectral modalities. These dependencies have been well documented, for example, in studies of vibronic coupling~\citep{Vibronic-Coupling}.
To effectively model these dependencies, we concatenate the encoded patch sequences from all spectra: $\hat{\vp} = \vp_1' \| \cdots \| \vp_{|\mathcal{S}|}' \in \mathbb{R}^{(\sum_{i=1}^{|\mathcal{S}|} N_i) \times d}$\hat{\vp} = \vp_1' \| \cdots \| \vp_{|\mathcal{S}|}|\mathcal{S}|' \in \mathbb{R}^{(\sum_{i=1}^{|\mathcal{S}|} N_i) \times d}(\sum_{i=1}i=1^{|\mathcal{S}|}|\mathcal{S}| N_i) \times d. This combined sequence is then passed through a Transformer encoder (see \cref{fig:overview}). In each attention head $h = 1, \ldots, H$h = 1, \ldots, H, queries, keys, and values are computed as $\bQ_h = \hat{\vp} \bW_h^Q$\bQ_h = \hat{\vp} \bW_h^Q, $\bK_h = \hat{\vp} \bW_h^K$\bK_h = \hat{\vp} \bW_h^K, and $\bV_h = \hat{\vp} \bW_h^V$\bV_h = \hat{\vp} \bW_h^V respectively, with $\bW_h^Q, \bW_h^K \in \mathbb{R}^{d \times d_k}$\bW_h^Q, \bW_h^K \in \mathbb{R}^{d \times d_k}d \times d_k and $\bW_h^V \in \mathbb{R}^{d \times \frac{d}{H}}$\bW_h^V \in \mathbb{R}^{d \times \frac{d}{H}}d \times \frac{d}{H}. The attention output for each head is:
\begin{equation}
\bO_h = \text{Attention}(\bQ_h, \bK_h, \bV_h) = \text{Softmax}\left(\frac{\bQ_h \bK_h^{\top}}{\sqrt{d_k}}\right)\bV_h.
\end{equation}\begin{equation}
\bO_h = \text{Attention}(\bQ_h, \bK_h, \bV_h) = \text{Softmax}\left(\frac{\bQ_h \bK_h^{\top}}{\sqrt{d_k}}\right)\bV_h.
\end{equation}
\bO_h = \text{Attention}(\bQ_h, \bK_h, \bV_h) = \text{Softmax}\left(\frac{\bQ_h \bK_h^{\top}}{\sqrt{d_k}}\right)\bV_h.
Batch normalization, feed-forward networks, and residual connections are integrated as illustrated in \cref{fig:overview}. The outputs from all attention heads are combined to form $\vz \in \mathbb{R}^{(\sum_{i=1}^{|\mathcal{S}|} N_i) \times d}$\vz \in \mathbb{R}^{(\sum_{i=1}^{|\mathcal{S}|} N_i) \times d}(\sum_{i=1}i=1^{|\mathcal{S}|}|\mathcal{S}| N_i) \times d. Finally, a flattening layer and a projection head are applied to produce the molecular spectra representation $\vz_s \in \mathbb{R}^d$\vz_s \in \mathbb{R}^d.
\subsubsection{Pre-training of SpecFormer}
To enable more effective encoding of molecular spectra, we introduce a masked reconstruction pre-training objective. In addition, to overcome the scarcity of spectral pre-training data by leveraging large-scale molecular structure pre-training, we incorporate a contrastive learning objective to align spectral and structural representations. The complete pre-training framework is illustrated in \cref{fig:overview}(C).
\textbf{Masked patches reconstruction pre-training for spectra.}
To ensure that the spectrum encoder can effectively extract and represent information from molecular spectra, we adopt a masked patches reconstruction (MPR) pre-training strategy. Inspired by the effectiveness of masked reconstruction across multiple domains~\citep{Bert,MAE,GraphMAE,Mole-Bert,AUG-MAE,PatchTST}, MPR guides the learning of SpecFormer by encouraging it to reconstruct masked portions of spectral data.
After segmenting the spectra into patches, we randomly mask a fraction of these patches—according to a predefined ratio $\alpha$\alpha—by replacing them with zero vectors. The masked patch sequences then undergo patch and position encoding, which conceals their semantic content (such as absorption intensities at certain wavelengths) but retains their positional information, aiding the reconstruction task.
Once processed by SpecFormer, the encoded representations corresponding to the masked patches are passed through a reconstruction head specific to each spectrum. The original values of the masked patches are then predicted, with the mean squared error (MSE) between the reconstructed and true patch values serving as the training objective:
\begin{equation}
\begin{aligned}
\mathcal{L}_{\mathrm{MPR}} = \sum_{i=1}^{|\mathcal{S}|} \mathbb{E}_{p_{i,j} \in \widetilde{\mathcal{P}}_i} \|\hat{\vp}_{i,j}- \vp_{i,j}\|_2^2,
\end{aligned}
\label{sce}
\end{equation}\begin{equation}
\begin{aligned}
\mathcal{L}_{\mathrm{MPR}} = \sum_{i=1}^{|\mathcal{S}|} \mathbb{E}_{p_{i,j} \in \widetilde{\mathcal{P}}_i} \|\hat{\vp}_{i,j}- \vp_{i,j}\|_2^2,
\end{aligned}
\label{sce}
\end{equation}
\mathcal{L}_{\mathrm{MPR}}\mathrm{MPR} = \sum_{i=1}i=1^{|\mathcal{S}|}|\mathcal{S}| \mathbb{E}_{p_{i,j} \in \widetilde{\mathcal{P}}_i}p_{i,j}i,j \in \widetilde{\mathcal{P}}_i \|\hat{\vp}_{i,j}i,j- \vp_{i,j}i,j\|_2^2,
\label{sce}
where $\widetilde{\mathcal{P}}_i$\widetilde{\mathcal{P}}_i is the set of masked patches for the $i$i-th type of spectrum, and $\hat{\vp}_{i,j}$\hat{\vp}_{i,j}i,j denotes the reconstruction for the masked patch $\vp_{i,j}$\vp_{i,j}i,j.
\textbf{Contrastive learning between 3D structures and spectra.}
To align spectral and 3D molecular representations, we introduce a contrastive learning objective in addition to the MPR. The 3D embeddings are learned under the guidance of a denoising objective~\cite{Coord,MolSpectra}. Here, the 3D embedding $\vz_x \in \mathbb{R}^d$\vz_x \in \mathbb{R}^d and spectra embedding $\vz_s \in \mathbb{R}^d$\vz_s \in \mathbb{R}^d of the same molecule are treated as positive pairs, while all other pairings are considered negative. The contrastive objective is designed to maximize similarity between positive pairs while minimizing similarity with negatives, using the InfoNCE loss~\citep{InfoNCE}:
\begin{equation}
\begin{aligned}
% \resizebox{\textwidth}{!}{$
\mathcal{L}_{\text{Contrast}} = -\frac{1}{2} \mathbb{E}_{p(\vz_x, \vz_s)} [ \log \frac{\exp(f_x(\vz_x, \vz_s))}{\exp(f_x(\vz_x, \vz_s)) + \sum_j \exp(f_x(\vz_x^j, \vz_s))} \\
+ \log \frac{\exp(f_s(\vz_s, \vz_x))}{\exp(f_s(\vz_s, \vz_x)) + \sum_j \exp(f_s(\vz_s^j, \vz_x))} ],
% $}
\label{eq:contrast}
\end{aligned}
\end{equation}\begin{equation}
\begin{aligned}
% \resizebox{\textwidth}{!}{$
\mathcal{L}_{\text{Contrast}} = -\frac{1}{2} \mathbb{E}_{p(\vz_x, \vz_s)} [ \log \frac{\exp(f_x(\vz_x, \vz_s))}{\exp(f_x(\vz_x, \vz_s)) + \sum_j \exp(f_x(\vz_x^j, \vz_s))} \\
+ \log \frac{\exp(f_s(\vz_s, \vz_x))}{\exp(f_s(\vz_s, \vz_x)) + \sum_j \exp(f_s(\vz_s^j, \vz_x))} ],
% $}
\label{eq:contrast}
\end{aligned}
\end{equation}
\mathcal{L}_{\text{Contrast}}\text{Contrast} = -\frac{1}{2} \mathbb{E}_{p(\vz_x, \vz_s)}p(\vz_x, \vz_s) [ \log \frac{\exp(f_x(\vz_x, \vz_s))}{\exp(f_x(\vz_x, \vz_s)) + \sum_j \exp(f_x(\vz_x^j, \vz_s))} \\
+ \log \frac{\exp(f_s(\vz_s, \vz_x))}{\exp(f_s(\vz_s, \vz_x)) + \sum_j \exp(f_s(\vz_s^j, \vz_x))} ],
\label{eq:contrast}
where $\vz_x^j$\vz_x^j and $\vz_s^j$\vz_s^j denote negative samples, and $f_x(\vz_x, \vz_s)$f_x(\vz_x, \vz_s) and $f_s(\vz_s, \vz_x)$f_s(\vz_s, \vz_x) are scoring functions, implemented here as the inner product: $f_x(\vz_x, \vz_s) = f_s(\vz_s, \vz_x) = \langle \vz_x, \vz_s \rangle$f_x(\vz_x, \vz_s) = f_s(\vz_s, \vz_x) = \langle \vz_x, \vz_s \rangle.
\textbf{Two-stage pre-training pipeline.}
Although spectral datasets are scarce, there exists a wealth of large-scale unlabeled molecular structure datasets. By leveraging our proposed contrastive alignment framework between spectra and structures, we can transfer knowledge from large-scale pre-training on molecular structures to enhance the learning of SpecFormer.
To fully exploit both spectral and structural information, we adopt a two-stage training protocol: the first stage performs pre-training on a large dataset~\citep{PCQM} containing only 3D structures using the denoising objective, while the second stage utilizes a dataset with available spectra to jointly optimize the overall objective:
\begin{equation}
\mathcal{L} = \beta_{\text{Denoising}} \mathcal{L}_{\text{Denoising}} + \beta_{\text{MPR}} \mathcal{L}_{\text{MPR}} + \beta_{\text{Contrast}} \mathcal{L}_{\text{Contrast}},
\label{eq:objective}
\end{equation}\begin{equation}
\mathcal{L} = \beta_{\text{Denoising}} \mathcal{L}_{\text{Denoising}} + \beta_{\text{MPR}} \mathcal{L}_{\text{MPR}} + \beta_{\text{Contrast}} \mathcal{L}_{\text{Contrast}},
\label{eq:objective}
\end{equation}
\mathcal{L} = \beta_{\text{Denoising}}\text{Denoising} \mathcal{L}_{\text{Denoising}}\text{Denoising} + \beta_{\text{MPR}}\text{MPR} \mathcal{L}_{\text{MPR}}\text{MPR} + \beta_{\text{Contrast}}\text{Contrast} \mathcal{L}_{\text{Contrast}}\text{Contrast},
\label{eq:objective}
where $\beta_{\text{Denoising}}$\beta_{\text{Denoising}}\text{Denoising}, $\beta_{\text{MPR}}$\beta_{\text{MPR}}\text{MPR}, and $\beta_{\text{Contrast}}$\beta_{\text{Contrast}}\text{Contrast} are weights for each component. In the second stage, SpecFormer is pre-trained exclusively on the QM9S training set to avoid data leakage.
\bibliography{main}
\clearpage
\captionsetup[figure]{labelformat=empty}labelformat=empty
\captionsetup[table]{labelformat=empty}labelformat=empty
\renewcommand{\thefigure}{Appendix Figure \arabic{figure}}
\renewcommand{\thetable}{Appendix Table \arabic{table}}
\begin{center}
\section*{\Large Appendix}
\end{center}\begin{center}
\section*{\Large Appendix}
\end{center}
|
Methods
| false
|
2507.06853
| 4
|
95,362
|
\vspace{2pt}
|
Appendix
| false
|
2507.06853
| 5
|
95,363
|
This appendix derives the reverse sampling formula for our continuous-time diffusion model, as applied to molecular graph generation. Our objective is to obtain an expression for sampling a less-noisy graph state $\cG_s$\cG_s from a more-noisy state $\cG_t = (\bH_t, \bA_t, \bX_t)$\cG_t = (\bH_t, \bA_t, \bX_t), given a model prediction of the original clean graph $\hat{\cG}_0 = (\hat{\bH}_0, \hat{\bA}_0, \hat{\bX}_0)$\hat{\cG}_0 = (\hat{\bH}_0, \hat{\bA}_0, \hat{\bX}_0). We assume a continuous time interval $[0, 1]$[0, 1], with $0 \leq s < t \leq 1$0 \leq s < t \leq 1.
To streamline the derivation, we consider a single graph component—either node features $\bH$\bH, edge features $\bA$\bA, or atomic coordinates $\bX$\bX—denoted generically as a vector-valued variable $\bG \in \mathbb{R}^d$\bG \in \mathbb{R}^d. The full derivation extends naturally by applying the result independently to each component.
\subsection{Forward Diffusion Process}
Following the formulation of VDM~\cite{VDM,ProgressiveDistillation,BlurringDiffusion}, the forward diffusion process is defined as:
\begin{equation}
q(\bG_t \mid \bG_0) = \mathcal{N}(\alpha_t \bG_0, \sigma_t^2 \bI),
\label{eq:forward}
\end{equation}\begin{equation}
q(\bG_t \mid \bG_0) = \mathcal{N}(\alpha_t \bG_0, \sigma_t^2 \bI),
\label{eq:forward}
\end{equation}
q(\bG_t \mid \bG_0) = \mathcal{N}(\alpha_t \bG_0, \sigma_t^2 \bI),
\label{eq:forward}
where the graph $\bG_t$\bG_t at time $t$t is a noisy version of the clean graph $\bG_0$\bG_0. Equivalently, it can be reparameterized as:
\begin{equation}
\bG_t = \alpha_t \bG_0 + \sigma_t \bG_\epsilon, \quad \bG_\epsilon \sim \mathcal{N}(\b0, \bI).
\end{equation}\begin{equation}
\bG_t = \alpha_t \bG_0 + \sigma_t \bG_\epsilon, \quad \bG_\epsilon \sim \mathcal{N}(\b0, \bI).
\end{equation}
\bG_t = \alpha_t \bG_0 + \sigma_t \bG_\epsilon, \quad \bG_\epsilon \sim \mathcal{N}(\b0, \bI).
Because $\alpha_t$\alpha_t decreases monotonically while $\sigma_t$\sigma_t increases, the information from $\bG_0$\bG_0 is gradually destroyed as $t$t increase. Assuming the process defined by~\cref{eq:forward} is Markovian, its transition distribution between two intermediate steps $s$s and $t$t with $0 \leq s < t$0 \leq s < t is:
\begin{equation}
q(\bG_t \mid \bG_s) = \mathcal{N}(\alpha_{t|s} \bG_s, \sigma_{t|s}^2 \bI),
\end{equation}\begin{equation}
q(\bG_t \mid \bG_s) = \mathcal{N}(\alpha_{t|s} \bG_s, \sigma_{t|s}^2 \bI),
\end{equation}
q(\bG_t \mid \bG_s) = \mathcal{N}(\alpha_{t|s}t|s \bG_s, \sigma_{t|s}t|s^2 \bI),
where $\alpha_{t|s} = \frac{\alpha_t}{\alpha_s}$\alpha_{t|s}t|s = \frac{\alpha_t}{\alpha_s} and $\sigma_{t|s}^2 = \sigma_t^2 - \alpha_{t|s}^2 \sigma_s^2$\sigma_{t|s}t|s^2 = \sigma_t^2 - \alpha_{t|s}t|s^2 \sigma_s^2. A convenient property of this framework is that the time grid can be defined arbitrarily and does not depend on the particular spacing of $s$s and $t$t. We set $T = 1$T = 1 to denote the final diffusion step, where $q(\bG_T \mid \bG_0) \approx \mathcal{N}(\b0, \bI)$q(\bG_T \mid \bG_0) \approx \mathcal{N}(\b0, \bI) approximates a standard normal distribution. Unless otherwise specified, time steps are assumed to lie in the unit interval $[0, 1]$[0, 1].
This formulation describes the distribution of a more-noised state $\bG_t$\bG_t conditioned on a less-noised state $\bG_s$\bG_s, and serves as a key component of the variational framework.
\subsection{Reverse Denoising Process Posterior $q(\bG_s \mid \bG_t, \bG_0)$}
We now derive the posterior distribution $q(\bG_s \mid \bG_t, \bG_0)$q(\bG_s \mid \bG_t, \bG_0), which describes the distribution of an intermediate state $\bG_s$\bG_s conditioned on both the noisy future $\bG_t$\bG_t and the original clean graph $\bG_0$\bG_0. This distribution underlies the training of the reverse process.
From standard Bayesian inference for Gaussians, suppose prior $\vz \sim \mathcal{N}(\boldsymbol{\mu}_p, \mathbf{\Sigma}_p)$\vz \sim \mathcal{N}(\boldsymbol{\mu}\mu_p, \mathbf{\Sigma}_p) and likelihood $\vy \mid \vz \sim \mathcal{N}(\bA \vz, \mathbf{\Sigma}_l)$\vy \mid \vz \sim \mathcal{N}(\bA \vz, \mathbf{\Sigma}_l). Then the posterior $p(\vz \mid \vy)$p(\vz \mid \vy) is Gaussian with:
\begin{gather}
\mathbf{\Sigma}_{\text{post}} = \left( \mathbf{\Sigma}_p^{-1} + \bA^\top \mathbf{\Sigma}_l^{-1} \bA \right)^{-1}, \\
\boldsymbol{\mu}_{\text{post}} = \mathbf{\Sigma}_{\text{post}} \left( \mathbf{\Sigma}_p^{-1} \boldsymbol{\mu}_p + \bA^\top \mathbf{\Sigma}_l^{-1} \vy \right).
\end{gather}\begin{gather}
\mathbf{\Sigma}_{\text{post}} = \left( \mathbf{\Sigma}_p^{-1} + \bA^\top \mathbf{\Sigma}_l^{-1} \bA \right)^{-1}, \\
\boldsymbol{\mu}_{\text{post}} = \mathbf{\Sigma}_{\text{post}} \left( \mathbf{\Sigma}_p^{-1} \boldsymbol{\mu}_p + \bA^\top \mathbf{\Sigma}_l^{-1} \vy \right).
\end{gather}
\mathbf{\Sigma}_{\text{post}}\text{post} = \left( \mathbf{\Sigma}_p^{-1}-1 + \bA^\top \mathbf{\Sigma}_l^{-1}-1 \bA \right)^{-1}-1, \\
\boldsymbol{\mu}\mu_{\text{post}}\text{post} = \mathbf{\Sigma}_{\text{post}}\text{post} \left( \mathbf{\Sigma}_p^{-1}-1 \boldsymbol{\mu}\mu_p + \bA^\top \mathbf{\Sigma}_l^{-1}-1 \vy \right).
In our case:
\begin{equation}
\begin{aligned}
\vz = \bG_s, \quad \vy &= \bG_t, \quad \boldsymbol{\mu}_p = \alpha_s \bG_0, \quad \mathbf{\Sigma}_p = \mathbf{\sigma}_s^2 \bI, \\
\bA &= \alpha_{t|s} \bI, \quad \mathbf{\Sigma}_l = \sigma_{t|s}^2 \bI.
\end{aligned}
\end{equation}\begin{equation}
\begin{aligned}
\vz = \bG_s, \quad \vy &= \bG_t, \quad \boldsymbol{\mu}_p = \alpha_s \bG_0, \quad \mathbf{\Sigma}_p = \mathbf{\sigma}_s^2 \bI, \\
\bA &= \alpha_{t|s} \bI, \quad \mathbf{\Sigma}_l = \sigma_{t|s}^2 \bI.
\end{aligned}
\end{equation}
\vz = \bG_s, \quad \vy &= \bG_t, \quad \boldsymbol{\mu}\mu_p = \alpha_s \bG_0, \quad \mathbf{\Sigma}_p = \mathbf{\sigma}_s^2 \bI, \\
\bA &= \alpha_{t|s}t|s \bI, \quad \mathbf{\Sigma}_l = \sigma_{t|s}t|s^2 \bI.
Applying this, we obtain:
\begin{equation}
q(\bG_s \mid \bG_t, \bG_0) = \mathcal{N}(\boldsymbol{\mu}_Q, \sigma_Q^2 \bI),
\end{equation}\begin{equation}
q(\bG_s \mid \bG_t, \bG_0) = \mathcal{N}(\boldsymbol{\mu}_Q, \sigma_Q^2 \bI),
\end{equation}
q(\bG_s \mid \bG_t, \bG_0) = \mathcal{N}(\boldsymbol{\mu}\mu_Q, \sigma_Q^2 \bI),
with:
\begin{gather}
\boldsymbol{\mu}_Q = \sigma_Q^2 \left( \frac{1}{\sigma_s^2} \alpha_s \bG_0 + \frac{\alpha_{t|s}}{\sigma_{t|s}^2} \bG_t \right)
= \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \bG_0, \\
\sigma_Q^2 = \left( \frac{1}{\sigma_s^2} + \frac{\alpha_{t|s}^2}{\sigma_{t|s}^2} \right)^{-1} = \frac{\sigma_s^2 \sigma_{t|s}^2}{\sigma_t^2}.
\end{gather}\begin{gather}
\boldsymbol{\mu}_Q = \sigma_Q^2 \left( \frac{1}{\sigma_s^2} \alpha_s \bG_0 + \frac{\alpha_{t|s}}{\sigma_{t|s}^2} \bG_t \right)
= \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \bG_0, \\
\sigma_Q^2 = \left( \frac{1}{\sigma_s^2} + \frac{\alpha_{t|s}^2}{\sigma_{t|s}^2} \right)^{-1} = \frac{\sigma_s^2 \sigma_{t|s}^2}{\sigma_t^2}.
\end{gather}
\boldsymbol{\mu}\mu_Q = \sigma_Q^2 \left( \frac{1}{\sigma_s^2} \alpha_s \bG_0 + \frac{\alpha_{t|s}}{\sigma_{t|s}^2} \bG_t \right)
= \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \bG_0, \\
\sigma_Q^2 = \left( \frac{1}{\sigma_s^2} + \frac{\alpha_{t|s}^2}{\sigma_{t|s}^2} \right)^{-1}-1 = \frac{\sigma_s^2 \sigma_{t|s}^2}{\sigma_t^2}.
The posterior mean is a convex combination of $\bG_t$\bG_t and $\bG_0$\bG_0, with weights determined by their respective noise scales. This provides the optimal denoised estimate of $\bG_s$\bG_s and forms the basis of the reverse sampling process.
\subsection{Reverse Denoising Process Approximation}
At inference time, the true $\bG_0$\bG_0 is unavailable, so we replace it with a model estimate $\hat{\bG}_0 = d_\theta(\bG_t, \tilde{\bG}_0, \log \mathrm{SNR}(t))$\hat{\bG}_0 = d_\theta(\bG_t, \tilde{\bG}_0, \log \mathrm{SNR}(t)). This yields the approximate reverse-time transition:
\begin{equation}
p(\bG_s \mid \bG_t) \approx \mathcal{N}(\bar{\bG}_s, \sigma_Q^2 \bI),
\end{equation}\begin{equation}
p(\bG_s \mid \bG_t) \approx \mathcal{N}(\bar{\bG}_s, \sigma_Q^2 \bI),
\end{equation}
p(\bG_s \mid \bG_t) \approx \mathcal{N}(\bar{\bG}_s, \sigma_Q^2 \bI),
where:
\begin{gather}
\bar{\bG}_s = \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \hat{\bG}_0, \\
\sigma_Q^2 = \frac{\sigma_s^2 \sigma_{t|s}^2}{\sigma_t^2}.
\end{gather}\begin{gather}
\bar{\bG}_s = \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \hat{\bG}_0, \\
\sigma_Q^2 = \frac{\sigma_s^2 \sigma_{t|s}^2}{\sigma_t^2}.
\end{gather}
\bar{\bG}_s = \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \hat{\bG}_0, \\
\sigma_Q^2 = \frac{\sigma_s^2 \sigma_{t|s}^2}{\sigma_t^2}.
We then sample from this distribution as:
\begin{equation}
\bG_s = \bar{\bG}_s + \frac{\sigma_s \sigma_{t|s}}{\sigma_t} \bG_\epsilon, \quad \bG_\epsilon \sim \mathcal{N}(\b0, \bI).
\end{equation}\begin{equation}
\bG_s = \bar{\bG}_s + \frac{\sigma_s \sigma_{t|s}}{\sigma_t} \bG_\epsilon, \quad \bG_\epsilon \sim \mathcal{N}(\b0, \bI).
\end{equation}
\bG_s = \bar{\bG}_s + \frac{\sigma_s \sigma_{t|s}}{\sigma_t} \bG_\epsilon, \quad \bG_\epsilon \sim \mathcal{N}(\b0, \bI).
Combining the above expressions, the complete reverse sampling step from $t$t to $s$s becomes:
\begin{equation}
\bG_s = \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \hat{\bG}_0 + \frac{\sigma_s \sigma_{t|s}}{\sigma_t} \bG_\epsilon.
\end{equation}\begin{equation}
\bG_s = \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \hat{\bG}_0 + \frac{\sigma_s \sigma_{t|s}}{\sigma_t} \bG_\epsilon.
\end{equation}
\bG_s = \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \hat{\bG}_0 + \frac{\sigma_s \sigma_{t|s}}{\sigma_t} \bG_\epsilon.
This closed-form expression enables efficient ancestral sampling in the reverse-time generative process.
\subsection{Sampling with Temperature}
To control the stochasticity in the sampling process, we introduce a temperature parameter $\tau > 0$\tau > 0, which scales the noise component during sampling:
\begin{equation}
\bG_s = \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \hat{\bG}_0 + \tau \cdot \frac{\sigma_s \sigma_{t|s}}{\sigma_t} \bG_\epsilon.
\end{equation}\begin{equation}
\bG_s = \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \hat{\bG}_0 + \tau \cdot \frac{\sigma_s \sigma_{t|s}}{\sigma_t} \bG_\epsilon.
\end{equation}
\bG_s = \frac{\alpha_{t|s} \sigma_s^2}{\sigma_t^2} \bG_t + \frac{\alpha_s \sigma_{t|s}^2}{\sigma_t^2} \hat{\bG}_0 + \tau \cdot \frac{\sigma_s \sigma_{t|s}}{\sigma_t} \bG_\epsilon.
The temperature parameter $\tau$\tau modulates the amount of stochasticity in the sampling process. When $\tau = 1$\tau = 1, the sampling follows the standard formulation. Lower values of $\tau < 1$\tau < 1 reduce the influence of noise, leading to more deterministic and potentially sharper outputs. Conversely, higher values of $\tau > 1$\tau > 1 increase the noise contribution, encouraging greater diversity at the cost of higher stochasticity. This flexibility allows practitioners to trade off between sample fidelity and variability based on downstream objectives.
|
Derivation of the Reverse Sampling Process
| false
|
2507.06853
| 6
|
95,364
|
We conduct experiments on the QM9S dataset~\cite{DetaNet,MolSpectra}, which augments the original QM9~\citep{QM9} molecular dataset with simulated spectra.
For each molecule, we extract its structure and corresponding IR, Raman, and UV-Vis spectra, forming a multi-modal spectral condition for our structure elucidation task.
The dataset contains 133{,},885 molecules in total.
Following the original setting in DetaNet~\citep{DetaNet}, we adopt the same training, validation, and test split strategy, which partitions the dataset into 90\% for training, 5\% for validation, and 5\% for testing.
|
Datasets
| false
|
2507.06853
| 7
|
95,365
|
Aiming to generating chemically valid and complete molecules, modeling accurate molecular distributions, and achieving achieve accurate sturcture elucidation from spectra, we elaborately design a series of evaluation metrics to reflect the generation quality.
Specifically, we introduce two categories of evaluation metrics. The first focuses on general molecular generation quality, emphasizing the validity and chemical soundness of the generated structures. The second targets molecular structure elucidation, evaluating the alignment between the predicted molecular structures—derived from spectral data—and the corresponding ground truth structures.
\subsection{Basic Metrics for Molecular Generation}
In the context of conditional molecule generation, these fundamental evaluation criteria serve as a prerequisite to guarantee that the generated structures are chemically meaningful and sufficiently diverse before evaluating their spectrum-conditioned structural accuracy.
\paragraph{Molecular Structure Validity Evaluation}Molecular Structure Validity Evaluation
The primary consideration is the chemical rationality of the generated molecular structures. The \textbf{Validity} metric ensures that the molecular structures conform to basic chemical rules, including chemically plausible valences (e.g., carbon atoms typically form four covalent bonds), and the overall bonding patterns are chemically reasonable (e.g., avoiding unrealistic bonds between atoms, maintaining aromatic delocalization). Building upon this, we introduce the \textbf{Validity and Complete (V\&C)} indicator, which not only requires chemical validity but also guarantees that the molecular graph forms a single connected entity instead of fragmented parts. This is particularly important for evaluating the model’s ability to generate fully connected, functional molecules.
To further assess the model’s capacity for innovation, we measure the proportions of molecules satisfying \textbf{Validity and Unique (V\&U)} and \textbf{Validity, Unique, and Novelty (V\&U\&N)} criteria. Uniqueness is determined by canonicalizing molecular representations and removing duplicates, thus ensuring diversity among the generated structures. Novelty is evaluated by comparing the generated molecules against the training set, reflecting the model’s ability to produce genuinely novel chemotypes rather than simply memorizing training data.
\paragraph{Molecular Structure Stability Evaluation}Molecular Structure Stability Evaluation
While structural validity is a necessary condition, it may not fully capture the chemical plausibility of a molecule. We additionally introduce a more stringent indicator of molecular stability. \textbf{Atom Stability} assesses whether each atom in the molecule has achieved an appropriate valence coordination, fully considering the influence of formal charges on the permissible bonding patterns. \textbf{Mol Stability} further requires that all atoms in the entire molecule satisfy the stability criteria, thereby providing a more comprehensive assessment of the chemical soundness of the generated molecules.
\paragraph{Distribution-based Structure Evaluation}Distribution-based Structure Evaluation
To evaluate the model's ability to learn the data distribution of real molecules, we adopt several distribution-based metrics. The Fréchet ChemNet Distance (\textbf{FCD}) measures the similarity between the distributions of generated molecules and reference molecules in a high-dimensional learned feature space. The \textbf{Similarity to Nearest Neighbor (SNN)} metric assesses the representativeness of the generated set by measuring the Tanimoto similarity between each generated molecule and its nearest neighbors in the test set, reflecting both diversity and coverage.
We additionally employ metrics based on molecular structural features: \textbf{Fragment Similarity (Frag)}, based on BRICS decomposition, evaluates distributional alignment at the functional group level; while \textbf{Scaffold Similarity (Scaf)}, using Bemis–Murcko scaffold analysis, quantifies similarity and diversity at the core-structure level. These complementary metrics jointly reflect the model’s ability to capture both local functional groups and global structural patterns, thereby characterizing the complexity of the generated molecular structures.
\paragraph{3D Geometry-based Structural Evaluation}3D Geometry-based Structural Evaluation
Beyond chemical rules in 2D representations, assessing the 3D geometric consistency of generated molecules is critical, especially for downstream applications such as molecular docking or molecular structure elucidation from spectra. We therefore introduce a set of geometric evaluation metrics based on fundamental molecular geometry features: bond lengths, bond angles, and dihedral angles.
\textbf{Bond} evaluation compares the distributions of bond lengths in the generated set against those in the reference set (test set molecules), capturing whether typical interatomic distances are preserved. \textbf{Angle} evaluation similarly assesses the distributions of bond angles, reflecting the local spatial arrangements of connected atoms and ring systems. \textbf{Dihedral} evaluation analyzes the distributions of torsional (dihedral) angles, providing insights into conformational consistency and stereochemical plausibility of the generated molecules.
To quantify the similarity between these geometric features of the generated and reference molecules, we employ maximum mean discrepancy (MMD), which measures the distance between two distributions in a kernel space. A lower MMD indicates that the generated molecules more faithfully reproduce realistic 3D geometries.
It is important to note that for all distribution-based and geometric metrics, the test set serves as the reference, while the novelty metric uses the training set as the reference to measure whether generated molecules are unseen. This distinction ensures a clear separation between distributional fidelity and generative innovation.
\subsection{Metrics for Molecular Structure Elucidation}\label{sec:metric2}
In the task of molecular structure elucidation, we introduce a set of precise metrics to evaluate the similarity between the ground-truth target structures and the generated candidate molecules under spectrum-conditioned generation. Here, a molecular structure is represented as a graph $\mathcal{G} = (\mathbf{H}, \mathbf{A}, \mathbf{X})$\mathcal{G} = (\mathbf{H}, \mathbf{A}, \mathbf{X}), where $\mathbf{H} \in \mathbb{R}^{N \times d_1}$\mathbf{H} \in \mathbb{R}^{N \times d_1}N \times d_1 denotes node-level attributes such as atom types and charges, $\mathbf{A} \in \mathbb{R}^{N \times N \times d_2}$\mathbf{A} \in \mathbb{R}^{N \times N \times d_2}N \times N \times d_2 encodes pairwise edge features such as bond existence and bond types, and $\mathbf{X} \in \mathbb{R}^{N \times 3}$\mathbf{X} \in \mathbb{R}^{N \times 3}N \times 3 corresponds to the 3D coordinates of the atoms.
\paragraph{Top-$K$ Accuracy}Top-$K$K Accuracy
The \textbf{Top-$K$ Accuracy} quantifies the model’s ability to recover the exact target structure among its top-$K$K generated candidates. For each condition characterized by a set of spectra, the model is sampled $K$K times to generate $K$K molecular structure candidates. If any of the $K$K candidates exactly matches the ground-truth structure, the prediction is considered correct. Formally, the metric is defined as:
\begin{equation}
\operatorname{ACC@}K = \mathbb{E}_{\mathcal{G}} \left[ \mathbbm{1} \left( \exists\, i \in \{1, \ldots, K\}, \; \hat{\mathcal{G}}_i = \mathcal{G} \right) \right],
\end{equation}\begin{equation}
\operatorname{ACC@}K = \mathbb{E}_{\mathcal{G}} \left[ \mathbbm{1} \left( \exists\, i \in \{1, \ldots, K\}, \; \hat{\mathcal{G}}_i = \mathcal{G} \right) \right],
\end{equation}
\operatorname{ACC@}ACC@K = \mathbb{E}_{\mathcal{G}}\mathcal{G} \left[ \mathbbm{1}1 \left( \exists\, i \in \{1, \ldots, K\}, \; \hat{\mathcal{G}}_i = \mathcal{G} \right) \right],
where $\mathbbm{1}(\cdot)$\mathbbm{1}1(\cdot) is the indicator function, $\mathcal{G}$\mathcal{G} denotes the ground-truth molecular graph, and $\hat{\mathcal{G}}_i$\hat{\mathcal{G}}_i denotes the $i$i-th generated candidate. This metric reflects the probability that the correct structure appears at least once among the top-$K$K model outputs.
\paragraph{Maximum Common Edge Subgraph (MCES)}Maximum Common Edge Subgraph (MCES)
To quantify graph-structural overlap, we adopt the \textbf{Maximum Common Edge Subgraph (MCES)} metric:
\begin{equation}
\operatorname{MCES} = \mathbb{E}_{\mathcal{G}} \left[
\max_{\mathcal{H} \subseteq \mathcal{G},\, \mathcal{H} \subseteq \hat{\mathcal{G}}} |E(\mathcal{H})|
\right],
\end{equation}\begin{equation}
\operatorname{MCES} = \mathbb{E}_{\mathcal{G}} \left[
\max_{\mathcal{H} \subseteq \mathcal{G},\, \mathcal{H} \subseteq \hat{\mathcal{G}}} |E(\mathcal{H})|
\right],
\end{equation}
\operatorname{MCES}MCES = \mathbb{E}_{\mathcal{G}}\mathcal{G} \left[
\max_{\mathcal{H} \subseteq \mathcal{G},\, \mathcal{H} \subseteq \hat{\mathcal{G}}}\mathcal{H} \subseteq \mathcal{G},\, \mathcal{H} \subseteq \hat{\mathcal{G}} |E(\mathcal{H})|
\right],
where $E(\mathcal{H})$E(\mathcal{H}) denotes the set of edges in the common subgraph $\mathcal{H}$\mathcal{H}. This indicator captures the extent of exact subgraph matching between the target and predicted molecular graphs, with higher values reflecting a greater degree of structural recovery.
\paragraph{Fingerprint-based Similarity Metrics}Fingerprint-based Similarity Metrics
We adopt several fingerprint-based similarity measures. Denote by $\mathbf{a}_{\mathrm{Morgan}}$\mathbf{a}_{\mathrm{Morgan}}\mathrm{Morgan} and $\hat{\mathbf{a}}_{\mathrm{Morgan}}$\hat{\mathbf{a}}_{\mathrm{Morgan}}\mathrm{Morgan} the binary Morgan fingerprint vectors of the target and predicted molecules, respectively, and by $\mathbf{a}_{\mathrm{MACCS}}$\mathbf{a}_{\mathrm{MACCS}}\mathrm{MACCS} and $\hat{\mathbf{a}}_{\mathrm{MACCS}}$\hat{\mathbf{a}}_{\mathrm{MACCS}}\mathrm{MACCS} their MACCS fingerprints. The \textbf{Tanimoto Similarity over Morgan fingerprints ($\operatorname{TaniSim}_{\mathrm{Morgan}}$)} is defined as:
\begin{equation}
\operatorname{TaniSim}_{\mathrm{Morgan}} =
\mathbb{E}_{\mathcal{G}} \left[
\frac{ \left| \mathbf{a}_{\mathrm{Morgan}} \land \hat{\mathbf{a}}_{\mathrm{Morgan}} \right| }
{ \left| \mathbf{a}_{\mathrm{Morgan}} \lor \hat{\mathbf{a}}_{\mathrm{Morgan}} \right| }
\right],
\end{equation}\begin{equation}
\operatorname{TaniSim}_{\mathrm{Morgan}} =
\mathbb{E}_{\mathcal{G}} \left[
\frac{ \left| \mathbf{a}_{\mathrm{Morgan}} \land \hat{\mathbf{a}}_{\mathrm{Morgan}} \right| }
{ \left| \mathbf{a}_{\mathrm{Morgan}} \lor \hat{\mathbf{a}}_{\mathrm{Morgan}} \right| }
\right],
\end{equation}
\operatorname{TaniSim}TaniSim_{\mathrm{Morgan}}\mathrm{Morgan} =
\mathbb{E}_{\mathcal{G}}\mathcal{G} \left[
\frac{ \left| \mathbf{a}_{\mathrm{Morgan}} \land \hat{\mathbf{a}}_{\mathrm{Morgan}} \right| }
{ \left| \mathbf{a}_{\mathrm{Morgan}} \lor \hat{\mathbf{a}}_{\mathrm{Morgan}} \right| }
\right],
where $\land$\land and $\lor$\lor denote bitwise AND and OR over the fingerprint vectors. This score quantifies the overlap in local structural patterns. The \textbf{Cosine Similarity over Morgan fingerprints ($\operatorname{CosSim}_{\mathrm{Morgan}}$)} is defined as:
\begin{equation}
\operatorname{CosSim}_{\mathrm{Morgan}} =
\mathbb{E}_{\mathcal{G}} \left[
\frac{ \mathbf{a}_{\mathrm{Morgan}} \cdot \hat{\mathbf{a}}_{\mathrm{Morgan}} }
{ \| \mathbf{a}_{\mathrm{Morgan}} \| \, \| \hat{\mathbf{a}}_{\mathrm{Morgan}} \| }
\right].
\end{equation}\begin{equation}
\operatorname{CosSim}_{\mathrm{Morgan}} =
\mathbb{E}_{\mathcal{G}} \left[
\frac{ \mathbf{a}_{\mathrm{Morgan}} \cdot \hat{\mathbf{a}}_{\mathrm{Morgan}} }
{ \| \mathbf{a}_{\mathrm{Morgan}} \| \, \| \hat{\mathbf{a}}_{\mathrm{Morgan}} \| }
\right].
\end{equation}
\operatorname{CosSim}CosSim_{\mathrm{Morgan}}\mathrm{Morgan} =
\mathbb{E}_{\mathcal{G}}\mathcal{G} \left[
\frac{ \mathbf{a}_{\mathrm{Morgan}} \cdot \hat{\mathbf{a}}_{\mathrm{Morgan}} }
{ \| \mathbf{a}_{\mathrm{Morgan}} \| \, \| \hat{\mathbf{a}}_{\mathrm{Morgan}} \| }
\right].
Likewise, the \textbf{Tanimoto Similarity over MACCS fingerprints ($\operatorname{TaniSim}_{\mathrm{MACCS}}$)} is:
\begin{equation}
\operatorname{TaniSim}_{\mathrm{MACCS}} =
\mathbb{E}_{\mathcal{G}} \left[
\frac{ \left| \mathbf{a}_{\mathrm{MACCS}} \land \hat{\mathbf{a}}_{\mathrm{MACCS}} \right| }
{ \left| \mathbf{a}_{\mathrm{MACCS}} \lor \hat{\mathbf{a}}_{\mathrm{MACCS}} \right| }
\right].
\end{equation}\begin{equation}
\operatorname{TaniSim}_{\mathrm{MACCS}} =
\mathbb{E}_{\mathcal{G}} \left[
\frac{ \left| \mathbf{a}_{\mathrm{MACCS}} \land \hat{\mathbf{a}}_{\mathrm{MACCS}} \right| }
{ \left| \mathbf{a}_{\mathrm{MACCS}} \lor \hat{\mathbf{a}}_{\mathrm{MACCS}} \right| }
\right].
\end{equation}
\operatorname{TaniSim}TaniSim_{\mathrm{MACCS}}\mathrm{MACCS} =
\mathbb{E}_{\mathcal{G}}\mathcal{G} \left[
\frac{ \left| \mathbf{a}_{\mathrm{MACCS}} \land \hat{\mathbf{a}}_{\mathrm{MACCS}} \right| }
{ \left| \mathbf{a}_{\mathrm{MACCS}} \lor \hat{\mathbf{a}}_{\mathrm{MACCS}} \right| }
\right].
These fingerprints capture interpretable functional patterns based on a set of predefined substructure keys, complementing the Morgan-based representations.
\paragraph{Fragment-based Similarity (Fraggle)}Fragment-based Similarity (Fraggle)
The \textbf{Fraggle Similarity} metric decomposes $\mathcal{G}$\mathcal{G} into a set of chemically meaningful fragments $\mathcal{F} = \{f_1, f_2, \dots, f_n\}$\mathcal{F} = \{f_1, f_2, \dots, f_n\} by strategies including double acyclic or exocyclic bond cuts with specified fragment size constraints. For each fragment $f_i$f_i, Fraggle computes two Tanimoto similarities over RDKit fingerprints between the target molecule $\mathcal{G}$\mathcal{G} and the predicted molecule $\hat{\mathcal{G}}$\hat{\mathcal{G}}: one based on the standard RDKit fingerprint over the entire molecule, and another based on a masked RDKit fingerprint where atoms outside $f_i$f_i with a Tversky similarity below 0.8 are replaced with wildcards. The fragment-level score is then taken as the maximum of these two values:
\begin{equation}
s_i = \max \left(
\operatorname{TaniSim}_{\mathrm{RDKit}}(\mathbf{a}_{\mathrm{RDKit}}, \hat{\mathbf{a}}_{\mathrm{RDKit}}),
\operatorname{TaniSim}_{\mathrm{RDKit}}(\mathbf{a}_{\mathrm{RDKit,mask}}, \hat{\mathbf{a}}_{\mathrm{RDKit,mask}})
\right).
\end{equation}\begin{equation}
s_i = \max \left(
\operatorname{TaniSim}_{\mathrm{RDKit}}(\mathbf{a}_{\mathrm{RDKit}}, \hat{\mathbf{a}}_{\mathrm{RDKit}}),
\operatorname{TaniSim}_{\mathrm{RDKit}}(\mathbf{a}_{\mathrm{RDKit,mask}}, \hat{\mathbf{a}}_{\mathrm{RDKit,mask}})
\right).
\end{equation}
s_i = \max \left(
\operatorname{TaniSim}TaniSim_{\mathrm{RDKit}}\mathrm{RDKit}(\mathbf{a}_{\mathrm{RDKit}}\mathrm{RDKit}, \hat{\mathbf{a}}_{\mathrm{RDKit}}\mathrm{RDKit}),
\operatorname{TaniSim}TaniSim_{\mathrm{RDKit}}\mathrm{RDKit}(\mathbf{a}_{\mathrm{RDKit,mask}}\mathrm{RDKit,mask}, \hat{\mathbf{a}}_{\mathrm{RDKit,mask}}\mathrm{RDKit,mask})
\right).
The final Fraggle similarity is defined as:
\begin{equation}
\operatorname{FraggleSim} = \mathbb{E}_{\mathcal{G}} \left[
\max_{f_i \in \mathcal{F}} s_i
\right].
\end{equation}\begin{equation}
\operatorname{FraggleSim} = \mathbb{E}_{\mathcal{G}} \left[
\max_{f_i \in \mathcal{F}} s_i
\right].
\end{equation}
\operatorname{FraggleSim}FraggleSim = \mathbb{E}_{\mathcal{G}}\mathcal{G} \left[
\max_{f_i \in \mathcal{F}}f_i \in \mathcal{F} s_i
\right].
This dual approach captures both global and local matching patterns
\paragraph{Functional Group-based Similarity}Functional Group-based Similarity
Lastly, to evaluate functional group consistency, we define a set of chemically significant functional groups (e.g., alkane, alcohol, amine, carboxylic acid, etc.) described by SMARTS patterns. For a pair of molecules $\mathcal{G}$\mathcal{G} (ground truth) and $\hat{\mathcal{G}}$\hat{\mathcal{G}} (prediction), we extract their respective functional group sets $FG(\mathcal{G})$FG(\mathcal{G}) and $FG(\hat{\mathcal{G}})$FG(\hat{\mathcal{G}}). The \textbf{Functional Group Similarity (FGSim)} is then computed as:
\begin{equation}
\operatorname{FGSim} =
\mathbb{E}_{\mathcal{G}} \left[
\frac{ \left| FG(\mathcal{G}) \cap FG(\hat{\mathcal{G}}) \right| }
{ \left| FG(\mathcal{G}) \cup FG(\hat{\mathcal{G}}) \right| }
\right].
\end{equation}\begin{equation}
\operatorname{FGSim} =
\mathbb{E}_{\mathcal{G}} \left[
\frac{ \left| FG(\mathcal{G}) \cap FG(\hat{\mathcal{G}}) \right| }
{ \left| FG(\mathcal{G}) \cup FG(\hat{\mathcal{G}}) \right| }
\right].
\end{equation}
\operatorname{FGSim}FGSim =
\mathbb{E}_{\mathcal{G}}\mathcal{G} \left[
\frac{ \left| FG(\mathcal{G}) \cap FG(\hat{\mathcal{G}}) \right| }
{ \left| FG(\mathcal{G}) \cup FG(\hat{\mathcal{G}}) \right| }
\right].
This interpretable metric is independent of molecular size, computationally efficient via substructure matching, and well-suited for comparative evaluation.
|
Metrics
| false
|
2507.06853
| 8
|
95,359
|
\label{sec:results}
\subsection{Overview of DiffSpectra Framework}
To achieve accurate molecular structure elucidation from molecular spectra, we propose \themodel, a spectrum-conditioned diffusion framework that generates molecular structures guided by spectral information. The overall paradigm is illustrated in \cref{fig:overview}.
Formally, a molecule characterized by its structure and spectra is represented as $\mathcal{M} = (\mathcal{G}, \mathcal{S})$\mathcal{M} = (\mathcal{G}, \mathcal{S}). The molecular structure is modeled as a graph $\mathcal{G} = (\mathbf{H}, \mathbf{A}, \mathbf{X})$\mathcal{G} = (\mathbf{H}, \mathbf{A}, \mathbf{X}), where $\mathbf{H} \in \mathbb{R}^{N \times d_1}$\mathbf{H} \in \mathbb{R}^{N \times d_1}N \times d_1 denotes node-level attributes such as atom types and charges, $\mathbf{A} \in \mathbb{R}^{N \times N \times d_2}$\mathbf{A} \in \mathbb{R}^{N \times N \times d_2}N \times N \times d_2 encodes pairwise edge features including bond existence and bond types, and $\mathbf{X} \in \mathbb{R}^{N \times 3}$\mathbf{X} \in \mathbb{R}^{N \times 3}N \times 3 corresponds to the 3D coordinates of the atoms. Here, $N$N is the number of atoms, $d_1$d_1 is the node feature dimensionality, and $d_2$d_2 is the edge feature dimensionality. The atomic ordering is consistent across $\mathbf{H}$\mathbf{H} and $\mathbf{X}$\mathbf{X} to preserve the correspondence between atom types and spatial positions.
Meanwhile, the molecular spectra are denoted as $\mathcal{S} = (\vs_1, \ldots, \vs_{|\mathcal{S}|})$\mathcal{S} = (\vs_1, \ldots, \vs_{|\mathcal{S}|}|\mathcal{S}|), where $|\mathcal{S}|=3$|\mathcal{S}|=3 in our setting, covering UV-Vis, IR, and Raman spectra. Concretely, $\vs_1 \in \mathbb{R}^{601}$\vs_1 \in \mathbb{R}^{601}601 records the UV-Vis spectrum from 1.5 to 13.5 eV with 0.02 eV intervals; $\vs_2 \in \mathbb{R}^{3501}$\vs_2 \in \mathbb{R}^{3501}3501 is the IR spectrum spanning 500–4000 cm$^{-1}$^{-1}-1 at 1 cm$^{-1}$^{-1}-1 resolution; and $\vs_3 \in \mathbb{R}^{3501}$\vs_3 \in \mathbb{R}^{3501}3501 is the Raman spectrum on the same range. Together, these modalities provide a comprehensive description of the molecular signatures.
\begin{figure}[H]
\begin{center}
\includegraphics[width=1\linewidth]{./figures/1_overview_v3.pdf}
\end{center}
\caption{(A) Overview of the \themodel framework, illustrating the continuous-time forward diffusion process and reverse denoising process. The Diffusion Molecule Transformer (DMT) is used as the denoising network, and spectra encoded by SpecFormer serve as conditional input.
(B) Architecture of the DMT, which processes node features, edge features, and atomic coordinates through three parallel streams with shared condition encoding, relational multi-head attention, and equivariant updates.
(C) Architecture and pre-training strategy of SpecFormer, which encodes multi-modal spectra (UV-Vis, IR, and Raman) through a unified transformer encoder. It is pre-trained via masked patch reconstruction (MPR) and contrastive learning with 3D molecular structures.}
\label{fig:overview}
\end{figure}
\includegraphics[width=1\linewidth]{./figures/1_overview_v3.pdf}
\caption{(A) Overview of the \themodel framework, illustrating the continuous-time forward diffusion process and reverse denoising process. The Diffusion Molecule Transformer (DMT) is used as the denoising network, and spectra encoded by SpecFormer serve as conditional input.
(B) Architecture of the DMT, which processes node features, edge features, and atomic coordinates through three parallel streams with shared condition encoding, relational multi-head attention, and equivariant updates.
(C) Architecture and pre-training strategy of SpecFormer, which encodes multi-modal spectra (UV-Vis, IR, and Raman) through a unified transformer encoder. It is pre-trained via masked patch reconstruction (MPR) and contrastive learning with 3D molecular structures.}
\label{fig:overview}
Within DiffSpectra, a continuous-time variational diffusion model progressively perturbs the molecular graph with Gaussian noise during the forward process, then denoises it in reverse, conditioned on spectra-derived information. A dedicated backbone, the Diffusion Molecule Transformer, parameterizes the denoising process while maintaining permutation and SE(3) equivariance. DMT is guided by spectral embeddings extracted by SpecFormer, a Transformer-based encoder pre-trained with both masked reconstruction and contrastive alignment to capture spectrum–structure correlations. This design allows the framework to achieve faithful and chemically meaningful structure elucidation, supported by physically grounded diffusion and rich spectral priors. More details of the full architecture and training objectives are presented in ~\cref{sec:method}.
\begin{table*}[h]
\centering
\caption{
Evaluation of basic metrics for molecular generation on the QM9S dataset.
Metrics include stability, validity, uniqueness, and novelty, as well as distribution-based and 3D geometry-based metrics.
The \textit{Train} row reports statistics from the training set as a reference distribution.
DiffSpectra is compared to unconditional molecular generative models.
}
\label{tab:validity}
\scriptsize
\begin{minipage}{\textwidth}
\centering
\makebox[\textwidth][c]{
% \resizebox{1.1\textwidth}{!}{
\begin{tabular}{lccccccccc}
\toprule
\textit{\textbf{Metric-2D}} & Atom stable $\uparrow$ & Mol stable $\uparrow$ & V\&C $\uparrow$ & V\&U $\uparrow$ & V\&U\&N $\uparrow$ & FCD $\downarrow$ & SNN $\uparrow$ & Frag $\uparrow$ & Scaf $\uparrow$ \\
\midrule
\textit{Train} & 99.9\% & 98.8\% & 98.9\% & 98.9\% & 0.0\% & 0.063 & 0.490 & 0.992 & 0.946 \\
\midrule
CDGS~\cite{CDGS} & 99.7\% & 95.1\% & 95.1\% & 93.6\% & 89.8\% & 0.798 & 0.493 & 0.973 & 0.784 \\
JODO~\cite{JODO} & \textbf{99.9\%} & \textbf{98.8\%} & \textbf{99.0\%} & {96.0\%} & 89.5\% & {0.138} & {0.522} & {0.986} & {0.934} \\
\midrule
\themodel & \textbf{99.9\%} & 98.2\% & 98.4\% & \textbf{96.8\%} & \textbf{93.7\%} & \textbf{0.088} & \textbf{0.531} & \textbf{0.992} & \textbf{0.943} \\
\bottomrule
\end{tabular}
% }
}
\end{minipage}
\par\vspace{3pt}
\begin{minipage}{\textwidth}
\centering
\makebox[\textwidth][c]{
% \resizebox{1.1\textwidth}{!}{
\begin{tabular}{lcccccc}
\toprule
\textit{\textbf{Metric-3D}} & Atom stable $\uparrow$ & Mol stable $\uparrow$ & FCD $\downarrow$ & Bond $\downarrow$ & Angle $\downarrow$ & Dihedral $\downarrow$ \\
\midrule
\textit{Train} & 99.4\% & 95.3\% & 0.877 & $5.44\text{e}{-4}$ & $4.65\text{e}{-4}$ & $1.78\text{e}{-4}$ \\
\midrule
E-NF~\cite{E-NF} & 84.7\% & 4.5\% & 4.452 & 0.6165 & 0.4203 & 0.0056 \\
G-SchNet~\cite{G-SchNet} & 95.7\% & 68.1\% & 2.386 & 0.3622 & 0.0727 & 0.0042 \\
G-SphereNet~\cite{G-SphereNet} & 67.2\% & 13.4\% & 6.659 & 0.1511 & 0.3537 & 0.0129 \\
EDM~\cite{EDM} & 98.6\% & 81.7\% & 1.285 & \textbf{0.1303} & 0.0182 & 6.64e-4 \\
MDM~\cite{MDM} & \textbf{99.2\%} & 89.6\% & 4.861 & 0.2735 & 0.0660 & 0.0239 \\
GeoLDM~\cite{GeoLDM} & 98.9\% & 89.7\% & 1.030 & 0.2404 & 0.0100 & 6.59e-4 \\
JODO~\cite{JODO} & \textbf{99.2\%} & \textbf{93.4\%} & \textbf{0.885} & 0.1475 & 0.0121 & 6.29e-4 \\
\midrule
\themodel & \textbf{99.2\%} & 92.9\% & 0.943 & 0.1792 & \textbf{0.0054} & \textbf{2.01e-4} \\
\bottomrule
\end{tabular}
% }
}
\end{minipage}
\end{table*}
\centering
\caption{
Evaluation of basic metrics for molecular generation on the QM9S dataset.
Metrics include stability, validity, uniqueness, and novelty, as well as distribution-based and 3D geometry-based metrics.
The \textit{Train} row reports statistics from the training set as a reference distribution.
DiffSpectra is compared to unconditional molecular generative models.
}
\label{tab:validity}
\scriptsize
{\textwidth}\textwidth
\centering
\makebox[\textwidth][c]{
% \resizebox{1.1\textwidth}{!}{
\begin{tabular}{lccccccccc}
\toprule
\textit{\textbf{Metric-2D}} & Atom stable $\uparrow$ & Mol stable $\uparrow$ & V\&C $\uparrow$ & V\&U $\uparrow$ & V\&U\&N $\uparrow$ & FCD $\downarrow$ & SNN $\uparrow$ & Frag $\uparrow$ & Scaf $\uparrow$ \\
\midrule
\textit{Train} & 99.9\% & 98.8\% & 98.9\% & 98.9\% & 0.0\% & 0.063 & 0.490 & 0.992 & 0.946 \\
\midrule
CDGS~\cite{CDGS} & 99.7\% & 95.1\% & 95.1\% & 93.6\% & 89.8\% & 0.798 & 0.493 & 0.973 & 0.784 \\
JODO~\cite{JODO} & \textbf{99.9\%} & \textbf{98.8\%} & \textbf{99.0\%} & {96.0\%} & 89.5\% & {0.138} & {0.522} & {0.986} & {0.934} \\
\midrule
\themodel & \textbf{99.9\%} & 98.2\% & 98.4\% & \textbf{96.8\%} & \textbf{93.7\%} & \textbf{0.088} & \textbf{0.531} & \textbf{0.992} & \textbf{0.943} \\
\bottomrule
\end{tabular}
% }
}
\toprule
\textit{\textbf{Metric-2D}} & Atom stable $\uparrow$\uparrow & Mol stable $\uparrow$\uparrow & V\&C $\uparrow$\uparrow & V\&U $\uparrow$\uparrow & V\&U\&N $\uparrow$\uparrow & FCD $\downarrow$\downarrow & SNN $\uparrow$\uparrow & Frag $\uparrow$\uparrow & Scaf $\uparrow$\uparrow \\
\midrule
\textit{Train} & 99.9\% & 98.8\% & 98.9\% & 98.9\% & 0.0\% & 0.063 & 0.490 & 0.992 & 0.946 \\
\midrule
CDGS~\cite{CDGS} & 99.7\% & 95.1\% & 95.1\% & 93.6\% & 89.8\% & 0.798 & 0.493 & 0.973 & 0.784 \\
JODO~\cite{JODO} & \textbf{99.9\%} & \textbf{98.8\%} & \textbf{99.0\%} & {96.0\%}96.0\% & 89.5\% & {0.138}0.138 & {0.522}0.522 & {0.986}0.986 & {0.934}0.934 \\
\midrule
\themodel & \textbf{99.9\%} & 98.2\% & 98.4\% & \textbf{96.8\%} & \textbf{93.7\%} & \textbf{0.088} & \textbf{0.531} & \textbf{0.992} & \textbf{0.943} \\
\bottomrule
\par\vspace{3pt}
{\textwidth}\textwidth
\centering
\makebox[\textwidth][c]{
% \resizebox{1.1\textwidth}{!}{
\begin{tabular}{lcccccc}
\toprule
\textit{\textbf{Metric-3D}} & Atom stable $\uparrow$ & Mol stable $\uparrow$ & FCD $\downarrow$ & Bond $\downarrow$ & Angle $\downarrow$ & Dihedral $\downarrow$ \\
\midrule
\textit{Train} & 99.4\% & 95.3\% & 0.877 & $5.44\text{e}{-4}$ & $4.65\text{e}{-4}$ & $1.78\text{e}{-4}$ \\
\midrule
E-NF~\cite{E-NF} & 84.7\% & 4.5\% & 4.452 & 0.6165 & 0.4203 & 0.0056 \\
G-SchNet~\cite{G-SchNet} & 95.7\% & 68.1\% & 2.386 & 0.3622 & 0.0727 & 0.0042 \\
G-SphereNet~\cite{G-SphereNet} & 67.2\% & 13.4\% & 6.659 & 0.1511 & 0.3537 & 0.0129 \\
EDM~\cite{EDM} & 98.6\% & 81.7\% & 1.285 & \textbf{0.1303} & 0.0182 & 6.64e-4 \\
MDM~\cite{MDM} & \textbf{99.2\%} & 89.6\% & 4.861 & 0.2735 & 0.0660 & 0.0239 \\
GeoLDM~\cite{GeoLDM} & 98.9\% & 89.7\% & 1.030 & 0.2404 & 0.0100 & 6.59e-4 \\
JODO~\cite{JODO} & \textbf{99.2\%} & \textbf{93.4\%} & \textbf{0.885} & 0.1475 & 0.0121 & 6.29e-4 \\
\midrule
\themodel & \textbf{99.2\%} & 92.9\% & 0.943 & 0.1792 & \textbf{0.0054} & \textbf{2.01e-4} \\
\bottomrule
\end{tabular}
% }
}
\toprule
\textit{\textbf{Metric-3D}} & Atom stable $\uparrow$\uparrow & Mol stable $\uparrow$\uparrow & FCD $\downarrow$\downarrow & Bond $\downarrow$\downarrow & Angle $\downarrow$\downarrow & Dihedral $\downarrow$\downarrow \\
\midrule
\textit{Train} & 99.4\% & 95.3\% & 0.877 & $5.44\text{e}{-4}$5.44\text{e}{-4}-4 & $4.65\text{e}{-4}$4.65\text{e}{-4}-4 & $1.78\text{e}{-4}$1.78\text{e}{-4}-4 \\
\midrule
E-NF~\cite{E-NF} & 84.7\% & 4.5\% & 4.452 & 0.6165 & 0.4203 & 0.0056 \\
G-SchNet~\cite{G-SchNet} & 95.7\% & 68.1\% & 2.386 & 0.3622 & 0.0727 & 0.0042 \\
G-SphereNet~\cite{G-SphereNet} & 67.2\% & 13.4\% & 6.659 & 0.1511 & 0.3537 & 0.0129 \\
EDM~\cite{EDM} & 98.6\% & 81.7\% & 1.285 & \textbf{0.1303} & 0.0182 & 6.64e-4 \\
MDM~\cite{MDM} & \textbf{99.2\%} & 89.6\% & 4.861 & 0.2735 & 0.0660 & 0.0239 \\
GeoLDM~\cite{GeoLDM} & 98.9\% & 89.7\% & 1.030 & 0.2404 & 0.0100 & 6.59e-4 \\
JODO~\cite{JODO} & \textbf{99.2\%} & \textbf{93.4\%} & \textbf{0.885} & 0.1475 & 0.0121 & 6.29e-4 \\
\midrule
\themodel & \textbf{99.2\%} & 92.9\% & 0.943 & 0.1792 & \textbf{0.0054} & \textbf{2.01e-4} \\
\bottomrule
\subsection{\themodel generates valid and stable molecular structures}
As summarized in \cref{tab:validity}, our proposed \themodel consistently achieves strong results in generating chemically valid and stable molecular structures under spectrum-conditioned settings. In the 2D evaluation, \themodel attains 99.9\% atom stability and 98.2\% molecular stability, demonstrating performance comparable to the training data distribution, while outperforming or matching general molecular generative models such as CDGS~\cite{CDGS} and JODO~\cite{JODO}. Moreover, \themodel maintains high structural uniqueness and novelty, with V\&U (valid and unique) at 96.8\% and V\&U\&N (valid, unique, and novel) at 93.7\%, significantly surpassing prior methods. In distribution-based metrics, \themodel achieves the lowest FCD, indicating a close distributional match to the reference test set. It also reaches the best SNN, Frag, and Scaf scores, demonstrating a strong capacity to cover diverse chemical substructures and scaffolds while maintaining validity.
In the 3D evaluation, \themodel demonstrates competitive geometric stability, with atom stability and molecular stability again comparable to the training set. Notably, \themodel achieves the best angle and dihedral MMDs, suggesting accurate recovery of bond angles and torsional angles, and achieves a bond distance MMD comparable to the top-performing baselines. This highlights \themodel’s effectiveness in reconstructing not only 2D connectivity but also realistic 3D molecular conformations.
The \textit{Train} row in \cref{tab:validity} reports the statistics of the training dataset as a reference distribution. For some metrics that measure similarity to the test set, such as SNN, Frag, Scaf, or geometric MMDs, the training data does not necessarily achieve the best possible scores, since the test distribution may differ from the training distribution. All other baselines listed in \cref{tab:validity}, including CDGS~\cite{CDGS}, JODO~\cite{JODO}, EDM~\cite{EDM}, and others, are unconditional molecular generative models, which sample from the general molecular space without explicit external guidance. In contrast, \themodel is specifically designed for molecular structure elucidation conditioned on spectra. The use of spectral signals as conditions provides external structural priors, enabling the model to generate results that better match the test set distribution. Therefore, the higher V\&U, V\&U\&N, SNN, Frag, Scaf, and MMD metrics achieved by \themodel compared to unconditional models are expected, as the conditioning naturally constrains the generation process toward structures consistent with the observed test samples.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{./figures/2_visualization_v2.pdf}
\end{center}
\caption{Visualization of molecular structure elucidation results using DiffSpectra under different configurations. We compare single-spectrum inputs (IR, Raman, UV-Vis), multi-modal spectra, and the effect of incorporating the pre-trained SpecFormer. Ground-truth structures are shown on the left for reference.}
\label{fig:visualization}
\end{figure}
\includegraphics[width=\linewidth]{./figures/2_visualization_v2.pdf}
\caption{Visualization of molecular structure elucidation results using DiffSpectra under different configurations. We compare single-spectrum inputs (IR, Raman, UV-Vis), multi-modal spectra, and the effect of incorporating the pre-trained SpecFormer. Ground-truth structures are shown on the left for reference.}
\label{fig:visualization}
\begin{table*}[h]
\centering
\caption{Structure elucidation performance of \themodel on the QM9S dataset.
We compare two configurations: one using a pre-trained SpecFormer as the spectral condition encoder, and one using an untrained SpecFormer.
Reported metrics include exact structure recovery, graph overlap, fingerprint-based similarities, fragment-level similarity, and functional group similarity.}
\label{tab:specformer}
\scriptsize
\makebox[\textwidth][c]{
\resizebox{1.05\textwidth}{!}{
\begin{tabular}{lcccccccc}
\toprule
\textit{\textbf{Model}}& \makecell{Pre-trained\\SpecFormer} & ACC@1 $\uparrow$ & MCES $\downarrow$ & $\operatorname{TaniSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{CosSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{TaniSim}_{\mathrm{MA}}$ $\uparrow$ & FraggleSim $\uparrow$ & FGSim $\uparrow$ \\
\midrule
\multirow{2}{*}{DiffSpectra} & \cmark & \first{16.01\%} & \first{1.3552} & \first{0.7837} & \first{0.8421} & \first{0.9227} & \first{0.9481} & \first{0.9618} \\
~ & \xmark & 14.11\% & 1.7795 & 0.7205 & 0.7938 & 0.8924 & 0.9383 & 0.9490 \\
\bottomrule
\end{tabular}
}
}
\end{table*}
\centering
\caption{Structure elucidation performance of \themodel on the QM9S dataset.
We compare two configurations: one using a pre-trained SpecFormer as the spectral condition encoder, and one using an untrained SpecFormer.
Reported metrics include exact structure recovery, graph overlap, fingerprint-based similarities, fragment-level similarity, and functional group similarity.}
\label{tab:specformer}
\scriptsize
\makebox[\textwidth][c]{
\resizebox{1.05\textwidth}{!}{
\begin{tabular}{lcccccccc}
\toprule
\textit{\textbf{Model}}& \makecell{Pre-trained\\SpecFormer} & ACC@1 $\uparrow$ & MCES $\downarrow$ & $\operatorname{TaniSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{CosSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{TaniSim}_{\mathrm{MA}}$ $\uparrow$ & FraggleSim $\uparrow$ & FGSim $\uparrow$ \\
\midrule
\multirow{2}{*}{DiffSpectra} & \cmark & \first{16.01\%} & \first{1.3552} & \first{0.7837} & \first{0.8421} & \first{0.9227} & \first{0.9481} & \first{0.9618} \\
~ & \xmark & 14.11\% & 1.7795 & 0.7205 & 0.7938 & 0.8924 & 0.9383 & 0.9490 \\
\bottomrule
\end{tabular}
}
}
\resizebox{1.05\textwidth}1.05\textwidth{!}!{
\begin{tabular}{lcccccccc}
\toprule
\textit{\textbf{Model}}& \makecell{Pre-trained\\SpecFormer} & ACC@1 $\uparrow$ & MCES $\downarrow$ & $\operatorname{TaniSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{CosSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{TaniSim}_{\mathrm{MA}}$ $\uparrow$ & FraggleSim $\uparrow$ & FGSim $\uparrow$ \\
\midrule
\multirow{2}{*}{DiffSpectra} & \cmark & \first{16.01\%} & \first{1.3552} & \first{0.7837} & \first{0.8421} & \first{0.9227} & \first{0.9481} & \first{0.9618} \\
~ & \xmark & 14.11\% & 1.7795 & 0.7205 & 0.7938 & 0.8924 & 0.9383 & 0.9490 \\
\bottomrule
\end{tabular}
}
\toprule
\textit{\textbf{Model}}& \makecell{Pre-trained\\SpecFormer}Pre-trained\\SpecFormer & ACC@1 $\uparrow$\uparrow & MCES $\downarrow$\downarrow & $\operatorname{TaniSim}_{\mathrm{MG}}$\operatorname{TaniSim}TaniSim_{\mathrm{MG}}\mathrm{MG} $\uparrow$\uparrow & $\operatorname{CosSim}_{\mathrm{MG}}$\operatorname{CosSim}CosSim_{\mathrm{MG}}\mathrm{MG} $\uparrow$\uparrow & $\operatorname{TaniSim}_{\mathrm{MA}}$\operatorname{TaniSim}TaniSim_{\mathrm{MA}}\mathrm{MA} $\uparrow$\uparrow & FraggleSim $\uparrow$\uparrow & FGSim $\uparrow$\uparrow \\
\midrule
\multirow{2}2{*}*{DiffSpectra}DiffSpectra & \cmark & \first{16.01\%}16.01\% & \first{1.3552}1.3552 & \first{0.7837}0.7837 & \first{0.8421}0.8421 & \first{0.9227}0.9227 & \first{0.9481}0.9481 & \first{0.9618}0.9618 \\
~ & \xmark & 14.11\% & 1.7795 & 0.7205 & 0.7938 & 0.8924 & 0.9383 & 0.9490 \\
\bottomrule
\subsection{\themodel accurately elucidates molecular structures from spectra}
As shown in \cref{tab:specformer} and \cref{fig:visualization}, \themodel demonstrates strong performance in the challenging task of molecular structure elucidation from spectra. First, the top-1 accuracy (ACC@1) reaches 16.01\%, meaning that the model is able to recover the exact target structure among its predictions in a nontrivial fraction of cases. Beyond strict exact matches, the average MCES score is 1.3552, indicating that even when the model fails to generate an exact structure, it still preserves a substantial portion of the molecular graph connectivity.
In terms of fingerprint-based similarity metrics, \themodel achieves a Tanimoto similarity of 0.7837 and a cosine similarity of 0.8421 over Morgan fingerprints, reflecting high agreement on local structural features. The Tanimoto similarity over MACCS keys is even higher at 0.9227, highlighting good coverage of predefined chemical substructures. These results suggest that \themodel captures the key functional groups and local motifs characteristic of the target molecules.
Furthermore, the fragment-based Fraggle similarity reaches 0.9481, showing that large chemically meaningful fragments are well recovered. The functional group similarity (FGSim) is also high at 0.9618, confirming that the predicted molecules retain nearly all functional group types present in the ground-truth structures. Overall, these metrics collectively indicate that \themodel is able to accurately elucidate molecular structures from spectra, recovering both global connectivity and local functional features with high fidelity.
\subsection{Pre-trained SpecFormer facilitates more accurate structure elucidation}
To investigate the effect of pre-training the spectrum encoder, we conduct an ablation study comparing DiffSpectra with and without a pre-trained SpecFormer module, as summarized in \cref{tab:specformer}. When equipped with the pre-trained SpecFormer, DiffSpectra achieves a top-1 accuracy of 16.01\%, compared to 14.11\% without pre-training, indicating a clear improvement in correctly recovering the ground-truth molecular structures from spectra. Similarly, the MCES decreases from 1.7795 to 1.3552, suggesting that the molecular graphs predicted with pre-trained SpecFormer retain more of the correct connectivity of the target structures.
In addition, the fingerprint-based similarity scores also improve with pre-training. For example, the Tanimoto similarity over Morgan fingerprints rises from 0.7205 to 0.7837, and cosine similarity improves from 0.7938 to 0.8421. The Tanimoto similarity over MACCS keys also increases from 0.8924 to 0.9227. These gains indicate that the molecular structures generated with pre-trained SpecFormer more accurately capture both local substructural patterns and functional group motifs. Moreover, the Fraggle similarity improves from 0.9383 to 0.9481, and FGSim from 0.9490 to 0.9618, confirming that larger chemical fragments and functional groups are better preserved.
Overall, these results demonstrate that pre-training the SpecFormer encoder on spectral data provides beneficial inductive biases, allowing DiffSpectra to more effectively align spectral representations with molecular graph structures during conditional generation. This highlights the value of leveraging a well-trained spectrum encoder to facilitate accurate molecular structure elucidation from spectra.
\begin{table*}[h]
\centering
\caption{Effect of spectral modalities on structure elucidation performance.
We report DiffSpectra's results on the QM9S dataset using different types of spectra as conditional input, including IR, Raman, UV-Vis, and their combination.
Metrics include exact structure recovery, graph overlap, fingerprint-based similarities, fragment-level similarity, and functional group similarity.}
\label{tab:modality}
\scriptsize
\makebox[\textwidth][c]{
\resizebox{1.05\textwidth}{!}{
\begin{tabular}{lcccccccc}
\toprule
\textit{\textbf{Model}} & Spectral Modalities & ACC@1 $\uparrow$ & MCES $\downarrow$ & $\operatorname{TaniSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{CosSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{TaniSim}_{\mathrm{MA}}$ $\uparrow$ & FraggleSim $\uparrow$ & FGSim $\uparrow$ \\
\midrule
\multirow{4}{*}{DiffSpectra} & All Spectra & \first{14.11\%} & \first{1.7795} & \first{0.7205} & \first{0.7938} & \first{0.8924} & \first{0.9383} & \first{0.9490} \\
~ & Only IR & 10.97\% & 2.4812 & 0.6246 & 0.7188 & 0.8460 & 0.9197 & 0.9269 \\
~ & Only Raman & 12.51\% & 2.1708 & 0.6778 & 0.7612 & 0.8701 & 0.9343 & 0.9315 \\
~ & Only UV-Vis & 0.10\% & 8.7909 & 0.1556 & 0.2625 & 0.3634 & 0.5581 & 0.4567 \\
\bottomrule
\end{tabular}
}
}
\end{table*}
\centering
\caption{Effect of spectral modalities on structure elucidation performance.
We report DiffSpectra's results on the QM9S dataset using different types of spectra as conditional input, including IR, Raman, UV-Vis, and their combination.
Metrics include exact structure recovery, graph overlap, fingerprint-based similarities, fragment-level similarity, and functional group similarity.}
\label{tab:modality}
\scriptsize
\makebox[\textwidth][c]{
\resizebox{1.05\textwidth}{!}{
\begin{tabular}{lcccccccc}
\toprule
\textit{\textbf{Model}} & Spectral Modalities & ACC@1 $\uparrow$ & MCES $\downarrow$ & $\operatorname{TaniSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{CosSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{TaniSim}_{\mathrm{MA}}$ $\uparrow$ & FraggleSim $\uparrow$ & FGSim $\uparrow$ \\
\midrule
\multirow{4}{*}{DiffSpectra} & All Spectra & \first{14.11\%} & \first{1.7795} & \first{0.7205} & \first{0.7938} & \first{0.8924} & \first{0.9383} & \first{0.9490} \\
~ & Only IR & 10.97\% & 2.4812 & 0.6246 & 0.7188 & 0.8460 & 0.9197 & 0.9269 \\
~ & Only Raman & 12.51\% & 2.1708 & 0.6778 & 0.7612 & 0.8701 & 0.9343 & 0.9315 \\
~ & Only UV-Vis & 0.10\% & 8.7909 & 0.1556 & 0.2625 & 0.3634 & 0.5581 & 0.4567 \\
\bottomrule
\end{tabular}
}
}
\resizebox{1.05\textwidth}1.05\textwidth{!}!{
\begin{tabular}{lcccccccc}
\toprule
\textit{\textbf{Model}} & Spectral Modalities & ACC@1 $\uparrow$ & MCES $\downarrow$ & $\operatorname{TaniSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{CosSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{TaniSim}_{\mathrm{MA}}$ $\uparrow$ & FraggleSim $\uparrow$ & FGSim $\uparrow$ \\
\midrule
\multirow{4}{*}{DiffSpectra} & All Spectra & \first{14.11\%} & \first{1.7795} & \first{0.7205} & \first{0.7938} & \first{0.8924} & \first{0.9383} & \first{0.9490} \\
~ & Only IR & 10.97\% & 2.4812 & 0.6246 & 0.7188 & 0.8460 & 0.9197 & 0.9269 \\
~ & Only Raman & 12.51\% & 2.1708 & 0.6778 & 0.7612 & 0.8701 & 0.9343 & 0.9315 \\
~ & Only UV-Vis & 0.10\% & 8.7909 & 0.1556 & 0.2625 & 0.3634 & 0.5581 & 0.4567 \\
\bottomrule
\end{tabular}
}
\toprule
\textit{\textbf{Model}} & Spectral Modalities & ACC@1 $\uparrow$\uparrow & MCES $\downarrow$\downarrow & $\operatorname{TaniSim}_{\mathrm{MG}}$\operatorname{TaniSim}TaniSim_{\mathrm{MG}}\mathrm{MG} $\uparrow$\uparrow & $\operatorname{CosSim}_{\mathrm{MG}}$\operatorname{CosSim}CosSim_{\mathrm{MG}}\mathrm{MG} $\uparrow$\uparrow & $\operatorname{TaniSim}_{\mathrm{MA}}$\operatorname{TaniSim}TaniSim_{\mathrm{MA}}\mathrm{MA} $\uparrow$\uparrow & FraggleSim $\uparrow$\uparrow & FGSim $\uparrow$\uparrow \\
\midrule
\multirow{4}4{*}*{DiffSpectra}DiffSpectra & All Spectra & \first{14.11\%}14.11\% & \first{1.7795}1.7795 & \first{0.7205}0.7205 & \first{0.7938}0.7938 & \first{0.8924}0.8924 & \first{0.9383}0.9383 & \first{0.9490}0.9490 \\
~ & Only IR & 10.97\% & 2.4812 & 0.6246 & 0.7188 & 0.8460 & 0.9197 & 0.9269 \\
~ & Only Raman & 12.51\% & 2.1708 & 0.6778 & 0.7612 & 0.8701 & 0.9343 & 0.9315 \\
~ & Only UV-Vis & 0.10\% & 8.7909 & 0.1556 & 0.2625 & 0.3634 & 0.5581 & 0.4567 \\
\bottomrule
\subsection{Multi-modal spectra outperform single-modality spectra for structure elucidation}
To further investigate the contribution of different spectral modalities, we conduct an ablation study where DiffSpectra is conditioned on individual or combined spectra types. As summarized in \cref{tab:modality}, using all spectra modalities (IR, Raman, UV-Vis) together achieves the best structure elucidation results, with a top-1 accuracy of 14.11\% and the lowest MCES of 1.7795, indicating improved graph connectivity recovery. The fingerprint-based metrics also show higher similarities when leveraging all spectra, with Tanimoto$_\text{MG}$_\text{MG} at 0.7205 and Cosine$_\text{MG}$_\text{MG} at 0.7938, suggesting better reconstruction of local molecular patterns. Likewise, the Tanimoto similarity on MACCS keys reaches 0.8924, and the Fraggle similarity is 0.9383, indicating strong preservation of meaningful chemical fragments.
Among individual modalities, Raman spectra alone outperform IR and UV-Vis, with ACC@1 of 12.51\% and reasonable similarity scores. IR spectra alone still provide useful chemical information, achieving 10.97\% top-1 accuracy, while UV-Vis alone performs poorly (0.1\% top-1 accuracy), reflecting its limited ability to uniquely identify molecular structures in the QM9S dataset. Overall, these results highlight that combining multiple spectra modalities provides complementary structural priors, enabling DiffSpectra to generate molecular structures with higher fidelity and consistency to ground-truth targets.
\subsection{Sampling multiple candidates improves structural hit accuracy}
\begin{wrapfigure}[22]{r}{0.5\linewidth}
\centering
\vspace{-1.3em}
\includegraphics[width=\linewidth]{./figures/4_sampling_times.pdf}
\vspace{-1em}
\caption{Accuracy@$K$ with increasing number of sampled candidates.
We evaluate the top-$K$ accuracy (ACC@$K$) as the number of generated candidates $K$ increases, comparing different variants of our model.
Across all settings, Accuracy@$K$ consistently improves with larger $K$, confirming that multiple sampling significantly increases the chance of recovering the correct molecular structure.}
\label{fig:sampling_times}
\end{wrapfigure}[22]{r}r{0.5\linewidth}0.5\linewidth
\centering
\vspace{-1.3em}
\includegraphics[width=\linewidth]{./figures/4_sampling_times.pdf}
\vspace{-1em}
\caption{Accuracy@$K$ with increasing number of sampled candidates.
We evaluate the top-$K$ accuracy (ACC@$K$) as the number of generated candidates $K$ increases, comparing different variants of our model.
Across all settings, Accuracy@$K$ consistently improves with larger $K$, confirming that multiple sampling significantly increases the chance of recovering the correct molecular structure.}
\label{fig:sampling_times}
Due to the inherently stochastic nature of diffusion models, each sampling process can produce a distinct yet plausible molecular structure. While a single sample may not always match the ground-truth structure exactly, generating multiple candidates increases the likelihood that at least one of them will align with the correct structure. This motivates the use of Top-$K$K Accuracy (ACC@$K$K) as a more comprehensive evaluation metric, which is formally defined in \cref{sec:metric2}.
We present in \cref{fig:sampling_times} the performance of different model variants under varying numbers of samples. Specifically, ACC@$K$K measures the proportion of test molecules for which the correct structure appears among the top $K$K generated candidates.
For instance, the ACC@1 reflects the exact match rate under single-sample generation, while ACC@$K$K ($K>1$K>1) demonstrate that allowing multiple guesses significantly boosts hit probability.
Across all model variants, we observe a consistent upward trend in Accuracy@$K$K as $K$K increases.
As $K$K increases, the model has more opportunities to generate a correct structure among its top-$K$K outputs. For example, the full model with all spectra and pre-trained SpecFormer improves from 16.01\% at $K=1$K=1 to 96.86\% at $K=20$K=20, showing that even a small sampling budget can significantly improve overall accuracy.
These results highlight an important property of diffusion-based molecular elucidation: even when the model may not always generate the correct structure in one shot, it maintains a strong ability to place the correct structure within a small number of plausible candidates. This aligns well with practical use cases, where a ranked list of likely structures can be provided for downstream validation or experimental testing.
The effectiveness of multi-sample inference further demonstrates how spectral conditioning constrains the generative space, enabling the model to consistently place the correct structure among top candidates.
\begin{table*}[h]
\centering
\caption{Evaluation of different SE(3) equivariance strategies for DiffSpectra.
We compare a model-based equivariant architecture to data-based equivariance approaches, with or without data augmentation, on molecular structure elucidation metrics. Here, ``data aug'' refers to data augmentation.
}
\label{tab:equivariance}
\scriptsize
\makebox[\textwidth][c]{
\resizebox{1.05\textwidth}{!}{
\begin{tabular}{lcccccccc}
\toprule
\textit{\textbf{Model}} & SE(3) Equivariance & ACC@1 $\uparrow$ & MCES $\downarrow$ & $\operatorname{TaniSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{CosSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{TaniSim}_{\mathrm{MA}}$ $\uparrow$ & FraggleSim $\uparrow$ & FGSim $\uparrow$ \\
\midrule
\multirow{3}{*}{DiffSpectra} & model-based & \first{14.11\%} & \first{1.7795} & 0.7205 & \first{0.7938} & \first{0.8924} & 0.9383 & 0.9490 \\
~ & data-based (w/ data aug) & 12.98\% & 1.8785 & \first{0.8882} & 0.7877 & 0.8882 & \first{0.9389} & \first{0.9456} \\
~ & data-based (w/o data aug) & 7.47\% & 3.6036 & 0.5117 & 0.6294 & 0.7575 & 0.8665 & 0.8607 \\
\bottomrule
\end{tabular}
}
}
\end{table*}
\centering
\caption{Evaluation of different SE(3) equivariance strategies for DiffSpectra.
We compare a model-based equivariant architecture to data-based equivariance approaches, with or without data augmentation, on molecular structure elucidation metrics. Here, ``data aug'' refers to data augmentation.
}
\label{tab:equivariance}
\scriptsize
\makebox[\textwidth][c]{
\resizebox{1.05\textwidth}{!}{
\begin{tabular}{lcccccccc}
\toprule
\textit{\textbf{Model}} & SE(3) Equivariance & ACC@1 $\uparrow$ & MCES $\downarrow$ & $\operatorname{TaniSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{CosSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{TaniSim}_{\mathrm{MA}}$ $\uparrow$ & FraggleSim $\uparrow$ & FGSim $\uparrow$ \\
\midrule
\multirow{3}{*}{DiffSpectra} & model-based & \first{14.11\%} & \first{1.7795} & 0.7205 & \first{0.7938} & \first{0.8924} & 0.9383 & 0.9490 \\
~ & data-based (w/ data aug) & 12.98\% & 1.8785 & \first{0.8882} & 0.7877 & 0.8882 & \first{0.9389} & \first{0.9456} \\
~ & data-based (w/o data aug) & 7.47\% & 3.6036 & 0.5117 & 0.6294 & 0.7575 & 0.8665 & 0.8607 \\
\bottomrule
\end{tabular}
}
}
\resizebox{1.05\textwidth}1.05\textwidth{!}!{
\begin{tabular}{lcccccccc}
\toprule
\textit{\textbf{Model}} & SE(3) Equivariance & ACC@1 $\uparrow$ & MCES $\downarrow$ & $\operatorname{TaniSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{CosSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{TaniSim}_{\mathrm{MA}}$ $\uparrow$ & FraggleSim $\uparrow$ & FGSim $\uparrow$ \\
\midrule
\multirow{3}{*}{DiffSpectra} & model-based & \first{14.11\%} & \first{1.7795} & 0.7205 & \first{0.7938} & \first{0.8924} & 0.9383 & 0.9490 \\
~ & data-based (w/ data aug) & 12.98\% & 1.8785 & \first{0.8882} & 0.7877 & 0.8882 & \first{0.9389} & \first{0.9456} \\
~ & data-based (w/o data aug) & 7.47\% & 3.6036 & 0.5117 & 0.6294 & 0.7575 & 0.8665 & 0.8607 \\
\bottomrule
\end{tabular}
}
\toprule
\textit{\textbf{Model}} & SE(3) Equivariance & ACC@1 $\uparrow$\uparrow & MCES $\downarrow$\downarrow & $\operatorname{TaniSim}_{\mathrm{MG}}$\operatorname{TaniSim}TaniSim_{\mathrm{MG}}\mathrm{MG} $\uparrow$\uparrow & $\operatorname{CosSim}_{\mathrm{MG}}$\operatorname{CosSim}CosSim_{\mathrm{MG}}\mathrm{MG} $\uparrow$\uparrow & $\operatorname{TaniSim}_{\mathrm{MA}}$\operatorname{TaniSim}TaniSim_{\mathrm{MA}}\mathrm{MA} $\uparrow$\uparrow & FraggleSim $\uparrow$\uparrow & FGSim $\uparrow$\uparrow \\
\midrule
\multirow{3}3{*}*{DiffSpectra}DiffSpectra & model-based & \first{14.11\%}14.11\% & \first{1.7795}1.7795 & 0.7205 & \first{0.7938}0.7938 & \first{0.8924}0.8924 & 0.9383 & 0.9490 \\
~ & data-based (w/ data aug) & 12.98\% & 1.8785 & \first{0.8882}0.8882 & 0.7877 & 0.8882 & \first{0.9389}0.9389 & \first{0.9456}0.9456 \\
~ & data-based (w/o data aug) & 7.47\% & 3.6036 & 0.5117 & 0.6294 & 0.7575 & 0.8665 & 0.8607 \\
\bottomrule
\subsection{Comparing model-based and data-based SE(3) equivariance strategies}
Since our model performs diffusion generation on molecular 3D structures, it is essential to ensure SE(3) equivariance so that the model's predictions are consistent under rigid-body transformations of the input. To this end, we design and evaluate two strategies for achieving SE(3) equivariance: model-based equivariance and data-based equivariance. In the model-based design, geometric inductive biases are explicitly incorporated into the network architecture, for example through pairwise distance encoding and equivariant coordinate update mechanisms, ensuring equivariant behavior. In contrast, the data-based design does not include SE(3)-equivariant inductive biases in the model architecture. Instead, it relies on random rotation and translation data augmentation during training, encouraging the model to empirically learn equivariance from data.
As shown in \cref{tab:equivariance}, the model-based SE(3)-equivariant implementation of DiffSpectra achieves superior performance across most metrics. It attains a top-1 accuracy of 14.11\% and the lowest MCES (1.7795), indicating stronger recovery of molecular graph connectivity. The data-based approach with data augmentation achieves slightly lower top-1 accuracy (12.98\%) but shows competitive fingerprint-based and functional group similarities, with Tanimoto$_{\mathrm{MG}}$_{\mathrm{MG}}\mathrm{MG}, FraggleSim, and FGSim comparable to the model-based design. Notably, when data augmentation is removed from the data-based design, performance degrades substantially across all metrics, confirming the critical role of data augmentation in enabling the model to generalize over molecular rotations and translations. Overall, these results demonstrate that while data-based equivariance with augmentation can help achieve some level of SE(3)-aware behavior, explicitly enforcing geometric inductive biases through model-based equivariance is a more robust and effective strategy for molecular structure elucidation from spectra.
\begin{table*}[h]
\centering
\caption{Effect of sampling temperature on structure elucidation performance.
We report DiffSpectra results with different sampling temperature coefficients $\tau$, which scale the injected noise during diffusion sampling.
Moderate values of $\tau$ help balance diversity and accuracy, while extremely low or high $\tau$ degrade performance.}
\label{tab:temperature}
\scriptsize
\makebox[\textwidth][c]{
\resizebox{1\textwidth}{!}{
\begin{tabular}{lcccccccc}
\toprule
\textit{\textbf{Model}} & Temperature & ACC@1 $\uparrow$ & MCES $\downarrow$ & $\operatorname{TaniSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{CosSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{TaniSim}_{\mathrm{MA}}$ $\uparrow$ & FraggleSim $\uparrow$ & FGSim $\uparrow$ \\
\midrule
\multirow{6}{*}{DiffSpectra} & $\tau$=1.2 & 15.45\% & 1.4224 & 0.7739 & 0.8348 & 0.9197 & 0.9468 & 0.9604 \\
~ & $\tau$=1.0 & 16.01\% & \first{1.3552} & 0.7837 & 0.8421 & 0.9227 & 0.9481 & 0.9618 \\
~ & $\tau$=0.8 & \first{16.30\%} & 1.3643 & \first{0.7848} & \first{0.8429} & \first{0.9235} & \first{0.9491} & 0.9621 \\
~ & $\tau$=0.6 & 16.05\% & 1.4009 & 0.7794 & 0.8387 & 0.9220 & 0.9480 & \first{0.9627} \\
~ & $\tau$=0.4 & 15.86\% & 1.5081 & 0.7629 & 0.8261 & 0.9120 & 0.9442 & 0.9577 \\
~ & $\tau$=0.2 & 14.56\% & 1.9842 & 0.7005 & 0.7774 & 0.8743 & 0.9297 & 0.9323 \\
\bottomrule
\end{tabular}
}
}
\end{table*}
\centering
\caption{Effect of sampling temperature on structure elucidation performance.
We report DiffSpectra results with different sampling temperature coefficients $\tau$, which scale the injected noise during diffusion sampling.
Moderate values of $\tau$ help balance diversity and accuracy, while extremely low or high $\tau$ degrade performance.}
\label{tab:temperature}
\scriptsize
\makebox[\textwidth][c]{
\resizebox{1\textwidth}{!}{
\begin{tabular}{lcccccccc}
\toprule
\textit{\textbf{Model}} & Temperature & ACC@1 $\uparrow$ & MCES $\downarrow$ & $\operatorname{TaniSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{CosSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{TaniSim}_{\mathrm{MA}}$ $\uparrow$ & FraggleSim $\uparrow$ & FGSim $\uparrow$ \\
\midrule
\multirow{6}{*}{DiffSpectra} & $\tau$=1.2 & 15.45\% & 1.4224 & 0.7739 & 0.8348 & 0.9197 & 0.9468 & 0.9604 \\
~ & $\tau$=1.0 & 16.01\% & \first{1.3552} & 0.7837 & 0.8421 & 0.9227 & 0.9481 & 0.9618 \\
~ & $\tau$=0.8 & \first{16.30\%} & 1.3643 & \first{0.7848} & \first{0.8429} & \first{0.9235} & \first{0.9491} & 0.9621 \\
~ & $\tau$=0.6 & 16.05\% & 1.4009 & 0.7794 & 0.8387 & 0.9220 & 0.9480 & \first{0.9627} \\
~ & $\tau$=0.4 & 15.86\% & 1.5081 & 0.7629 & 0.8261 & 0.9120 & 0.9442 & 0.9577 \\
~ & $\tau$=0.2 & 14.56\% & 1.9842 & 0.7005 & 0.7774 & 0.8743 & 0.9297 & 0.9323 \\
\bottomrule
\end{tabular}
}
}
\resizebox{1\textwidth}1\textwidth{!}!{
\begin{tabular}{lcccccccc}
\toprule
\textit{\textbf{Model}} & Temperature & ACC@1 $\uparrow$ & MCES $\downarrow$ & $\operatorname{TaniSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{CosSim}_{\mathrm{MG}}$ $\uparrow$ & $\operatorname{TaniSim}_{\mathrm{MA}}$ $\uparrow$ & FraggleSim $\uparrow$ & FGSim $\uparrow$ \\
\midrule
\multirow{6}{*}{DiffSpectra} & $\tau$=1.2 & 15.45\% & 1.4224 & 0.7739 & 0.8348 & 0.9197 & 0.9468 & 0.9604 \\
~ & $\tau$=1.0 & 16.01\% & \first{1.3552} & 0.7837 & 0.8421 & 0.9227 & 0.9481 & 0.9618 \\
~ & $\tau$=0.8 & \first{16.30\%} & 1.3643 & \first{0.7848} & \first{0.8429} & \first{0.9235} & \first{0.9491} & 0.9621 \\
~ & $\tau$=0.6 & 16.05\% & 1.4009 & 0.7794 & 0.8387 & 0.9220 & 0.9480 & \first{0.9627} \\
~ & $\tau$=0.4 & 15.86\% & 1.5081 & 0.7629 & 0.8261 & 0.9120 & 0.9442 & 0.9577 \\
~ & $\tau$=0.2 & 14.56\% & 1.9842 & 0.7005 & 0.7774 & 0.8743 & 0.9297 & 0.9323 \\
\bottomrule
\end{tabular}
}
\toprule
\textit{\textbf{Model}} & Temperature & ACC@1 $\uparrow$\uparrow & MCES $\downarrow$\downarrow & $\operatorname{TaniSim}_{\mathrm{MG}}$\operatorname{TaniSim}TaniSim_{\mathrm{MG}}\mathrm{MG} $\uparrow$\uparrow & $\operatorname{CosSim}_{\mathrm{MG}}$\operatorname{CosSim}CosSim_{\mathrm{MG}}\mathrm{MG} $\uparrow$\uparrow & $\operatorname{TaniSim}_{\mathrm{MA}}$\operatorname{TaniSim}TaniSim_{\mathrm{MA}}\mathrm{MA} $\uparrow$\uparrow & FraggleSim $\uparrow$\uparrow & FGSim $\uparrow$\uparrow \\
\midrule
\multirow{6}6{*}*{DiffSpectra}DiffSpectra & $\tau$\tau=1.2 & 15.45\% & 1.4224 & 0.7739 & 0.8348 & 0.9197 & 0.9468 & 0.9604 \\
~ & $\tau$\tau=1.0 & 16.01\% & \first{1.3552}1.3552 & 0.7837 & 0.8421 & 0.9227 & 0.9481 & 0.9618 \\
~ & $\tau$\tau=0.8 & \first{16.30\%}16.30\% & 1.3643 & \first{0.7848}0.7848 & \first{0.8429}0.8429 & \first{0.9235}0.9235 & \first{0.9491}0.9491 & 0.9621 \\
~ & $\tau$\tau=0.6 & 16.05\% & 1.4009 & 0.7794 & 0.8387 & 0.9220 & 0.9480 & \first{0.9627}0.9627 \\
~ & $\tau$\tau=0.4 & 15.86\% & 1.5081 & 0.7629 & 0.8261 & 0.9120 & 0.9442 & 0.9577 \\
~ & $\tau$\tau=0.2 & 14.56\% & 1.9842 & 0.7005 & 0.7774 & 0.8743 & 0.9297 & 0.9323 \\
\bottomrule
\subsection{Sampling temperature balances diversity and accuracy}
Since a given molecular spectrum corresponds to a unique molecular structure, the task of molecular structure elucidation from spectra requires high accuracy in the generated results. However, diffusion models inherently introduce stochasticity during sampling due to the injection of random noise, which is essential for promoting diversity in the generated results. To better understand the trade-off between stochasticity and accuracy, we investigate how controlling the level of stochasticity affects performance. Specifically, we introduce a sampling temperature parameter $\tau$\tau in \cref{eq:sampling}, which modulates the scale of the injected noise during the sampling process. Lower $\tau$\tau values reduce the influence of stochasticity, resulting in more deterministic outputs, while higher $\tau$\tau allows for greater exploration of the molecular space.
As shown in \cref{tab:temperature}, we evaluate DiffSpectra under varying $\tau$\tau values ranging from 0.2 to 1.2. We observe that moderate temperature values ($\tau=0.8$\tau=0.8 and $\tau=1.0$\tau=1.0) achieve the best balance between diversity and correctness. In particular, $\tau=0.8$\tau=0.8 obtains the highest top-1 accuracy of 16.30\%, while maintaining competitive similarity metrics, including a Tanimoto$_{\mathrm{MG}}$_{\mathrm{MG}}\mathrm{MG} of 0.7848, Cosine$_{\mathrm{MG}}$_{\mathrm{MG}}\mathrm{MG} of 0.8429, and Tanimoto$_{\mathrm{MA}}$_{\mathrm{MA}}\mathrm{MA} of 0.9235. Extremely low temperatures (e.g., $\tau=0.2$\tau=0.2) result in overly deterministic generations with reduced diversity and higher MCES (1.9842), while extremely high temperatures (e.g., $\tau=1.2$\tau=1.2) also degrade performance due to excessive stochasticity.
Overall, these results suggest that selecting a moderate $\tau$\tau helps DiffSpectra strike an effective trade-off between generating structurally diverse candidates and maintaining high similarity to the ground-truth molecules.
|
Results
| false
|
2507.06853
| 2
|
95,369
|
Now we want to define the notion of derivability for $R$R-expressions of arbitrary $R$R-degree.
Therefore, we define a calculus in which not only sequents but $R$R-expressions of arbitrary degree are allowed.
The following structural rules hold:
\begin{center}
\vspace{0.4cm}
\quad \infer[RF]{S\Rightarrow S}{}\hspace{7em}\infer[WL]{\Delta, T \Rightarrow S}{\Delta \Rightarrow S}
\vspace{0.2cm}
\quad\infer[PL]{\Delta, T, S, \Gamma \Rightarrow U}{\Delta, S, T, \Gamma \Rightarrow U}\hspace{6em} \infer[CL]{\Delta, S \Rightarrow T}{\Delta, S, S \Rightarrow T }
\vspace{0.2cm}
\quad \infer[Cut]{\Delta \Rightarrow T}{\Delta \Rightarrow S \qquad \Delta, S \Rightarrow T}
\end{center}\begin{center}
\vspace{0.4cm}
\quad \infer[RF]{S\Rightarrow S}{}\hspace{7em}\infer[WL]{\Delta, T \Rightarrow S}{\Delta \Rightarrow S}
\vspace{0.2cm}
\quad\infer[PL]{\Delta, T, S, \Gamma \Rightarrow U}{\Delta, S, T, \Gamma \Rightarrow U}\hspace{6em} \infer[CL]{\Delta, S \Rightarrow T}{\Delta, S, S \Rightarrow T }
\vspace{0.2cm}
\quad \infer[Cut]{\Delta \Rightarrow T}{\Delta \Rightarrow S \qquad \Delta, S \Rightarrow T}
\end{center}
\vspace{0.4cm}
\quad \infer[RF]{S\Rightarrow S}S\Rightarrow S{}\hspace{7em}\infer[WL]{\Delta, T \Rightarrow S}\Delta, T \Rightarrow S{\Delta \Rightarrow S}\Delta \Rightarrow S
\vspace{0.2cm}
\quad\infer[PL]{\Delta, T, S, \Gamma \Rightarrow U}\Delta, T, S, \Gamma \Rightarrow U{\Delta, S, T, \Gamma \Rightarrow U}\Delta, S, T, \Gamma \Rightarrow U\hspace{6em} \infer[CL]{\Delta, S \Rightarrow T}\Delta, S \Rightarrow T{\Delta, S, S \Rightarrow T }\Delta, S, S \Rightarrow T
\vspace{0.2cm}
\quad \infer[Cut]{\Delta \Rightarrow T}\Delta \Rightarrow T{\Delta \Rightarrow S \qquad \Delta, S \Rightarrow T}\Delta \Rightarrow S \qquad \Delta, S \Rightarrow T
\noindent
We add two rules, which tell us how to introduce an $R$R-expression of arbitrary degree on the right and on the left side of $\Rightarrow$\Rightarrow:
\begin{center}
\vspace{0.3cm}
\quad \infer[RI^+]{\Gamma \Rightarrow (\Delta \Rightarrow S)}{\Gamma, \Delta \Rightarrow S}\quad\quad \infer[LI^+]{ \Gamma, (\Delta \Rightarrow S) \Rightarrow T}{\Gamma \Rightarrow \Delta \quad\quad \Gamma, S \Rightarrow T}
\vspace{0.2cm}
\end{center}\begin{center}
\vspace{0.3cm}
\quad \infer[RI^+]{\Gamma \Rightarrow (\Delta \Rightarrow S)}{\Gamma, \Delta \Rightarrow S}\quad\quad \infer[LI^+]{ \Gamma, (\Delta \Rightarrow S) \Rightarrow T}{\Gamma \Rightarrow \Delta \quad\quad \Gamma, S \Rightarrow T}
\vspace{0.2cm}
\end{center}
\vspace{0.3cm}
\quad \infer[RI^+]{\Gamma \Rightarrow (\Delta \Rightarrow S)}\Gamma \Rightarrow (\Delta \Rightarrow S){\Gamma, \Delta \Rightarrow S}\Gamma, \Delta \Rightarrow S\quad\quad \infer[LI^+]{ \Gamma, (\Delta \Rightarrow S) \Rightarrow T} \Gamma, (\Delta \Rightarrow S) \Rightarrow T{\Gamma \Rightarrow \Delta \quad\quad \Gamma, S \Rightarrow T}\Gamma \Rightarrow \Delta \quad\quad \Gamma, S \Rightarrow T
\vspace{0.2cm}
\noindent The expression $\Gamma \Rightarrow \Delta$\Gamma \Rightarrow \Delta is here to be read as $\Gamma \Rightarrow U_1;...;\Gamma \Rightarrow U_n$\Gamma \Rightarrow U_1;...;\Gamma \Rightarrow U_n where $\Delta$\Delta is the series of $R$R-expressions $U_1,...U_n$U_1,...U_n.
\noindent In line with our connexive understanding of inference, we define the refutability of $\Gamma \Rightarrow S$\Gamma \Rightarrow S by the refutability of $S$S under the assumption of the provability of $\Gamma$\Gamma and the refutability of $\Gamma \Rightarrow -S$\Gamma \Rightarrow -S by the provability of $S$S under the assumption of the provability of $\Gamma$\Gamma.
To formulate this as one rule, we can fix that $--S$--S stands for $S$S.
Then we have the following rule:
\begin{center}
\vspace{0.3cm}
\quad\infer[RI^-]{ \Delta \Rightarrow -(\Gamma \Rightarrow S)}{\Delta, \Gamma \Rightarrow -S}
\vspace{0.2cm}
\end{center}\begin{center}
\vspace{0.3cm}
\quad\infer[RI^-]{ \Delta \Rightarrow -(\Gamma \Rightarrow S)}{\Delta, \Gamma \Rightarrow -S}
\vspace{0.2cm}
\end{center}
\vspace{0.3cm}
\quad\infer[RI^-]{ \Delta \Rightarrow -(\Gamma \Rightarrow S)} \Delta \Rightarrow -(\Gamma \Rightarrow S){\Delta, \Gamma \Rightarrow -S}\Delta, \Gamma \Rightarrow -S
\vspace{0.2cm}
\noindent If the reverse is demanded to hold as well, i.e.:
\begin{center}
\vspace{0.2cm}
\quad\infer{\Delta \Rightarrow -(\Gamma \Rightarrow S)}{ \Delta, \Gamma \Rightarrow -S}
\vspace{0.2cm}
\end{center}\begin{center}
\vspace{0.2cm}
\quad\infer{\Delta \Rightarrow -(\Gamma \Rightarrow S)}{ \Delta, \Gamma \Rightarrow -S}
\vspace{0.2cm}
\end{center}
\vspace{0.2cm}
\quad\infer{\Delta \Rightarrow -(\Gamma \Rightarrow S)}\Delta \Rightarrow -(\Gamma \Rightarrow S){ \Delta, \Gamma \Rightarrow -S} \Delta, \Gamma \Rightarrow -S
\vspace{0.2cm}
\noindent then we also obtain a condition on how to derive something \emph{from} $R$R-expressions of the form $-(\Gamma \Rightarrow S)$-(\Gamma \Rightarrow S):
\begin{center}
\vspace{0.2cm}
\quad\infer[LI^-]{\Delta, -(\Gamma \Rightarrow S) \Rightarrow T}{\Delta \Rightarrow \Gamma \quad\quad \Delta, -S \Rightarrow T}
\vspace{0.2cm}
\end{center}\begin{center}
\vspace{0.2cm}
\quad\infer[LI^-]{\Delta, -(\Gamma \Rightarrow S) \Rightarrow T}{\Delta \Rightarrow \Gamma \quad\quad \Delta, -S \Rightarrow T}
\vspace{0.2cm}
\end{center}
\vspace{0.2cm}
\quad\infer[LI^-]{\Delta, -(\Gamma \Rightarrow S) \Rightarrow T}\Delta, -(\Gamma \Rightarrow S) \Rightarrow T{\Delta \Rightarrow \Gamma \quad\quad \Delta, -S \Rightarrow T}\Delta \Rightarrow \Gamma \quad\quad \Delta, -S \Rightarrow T
\vspace{0.2cm}
We will write $\Delta\vdash S$\Delta\vdash S to mean that there is a derivation from the $R$R-expressions of $\Delta$\Delta to $S$S, and we write $\Delta\vdash \Gamma$\Delta\vdash \Gamma to mean that $\Delta\vdash S$\Delta\vdash S for each $S$S from $\Gamma$\Gamma.
Now that we have our deductive framework, let us briefly outline the remainder of the paper.
First, we adopt von Kutschera's generalized introduction rules for arbitrary connectives.
We then give the introduction rules for the usual connectives for \texttt{C} and follow von Kutschera's argument step by step to obtain functional completeness for \texttt{C}, meaning that any connective expressible by the generalized schemata can be explicitly defined through the connectives of \texttt{C}.
Finally, we provide a way to translate this framework into a standard sequent calculus of \texttt{C} and obtain that a sequent is derivable in this framework iff its translation is derivable in \texttt{C}.
|
Higher-Order Derivability
| false
|
2507.06854
| 4
|
95,366
|
From the viewpoint of what would nowadays be called `(bilateralist) proof-theoretic semantics' Franz von Kutschera provided two very relevant and closely related papers \cite{Kutschera1968,Kutschera1969}, in which he proves functional completeness of, respectively, intuitionistic logic and what he calls `direct logic', nowadays better known as Nelson's constructive logic with strong negation, \texttt{N4}.
Proof-theoretic semantics is the view that it is the rules of logical connectives governing their use in a proof system that determine their meaning.\footnote{Not unwarrantedly, von Kutschera called this `Gentzen semantics', since this view's origins are usually attributed to Gentzen's famous paper \cite{Gentzen1934}. For more details on proof-theoretic semantics, see \cite{s-h} and \cite{Francez2015}.}
Von Kutschera commits to this view in these papers in that he assumes rule schemata to define connectives and in conducting the proofs for functional completeness purely proof-theoretically, i.e., without reference to model-theoretic notions.
The procedure is in both papers essentially the same.
In a nutshell, he considers generalized derivability schemata for unspecified logical operators and shows that the connectives of intuitionistic propositional logic and \texttt{N4}, respectively, are functionally complete when defined according to these rule schemata.
The aim of the present paper is to transfer this approach to show the functional completeness for the connectives of the connexive logic \texttt{C}.
As such, our paper is in a tradition of several others using von Kutschera's approach in one way or another; to name just a few:
Schroeder-Heister, combining von Kutschera's higher level derivability with Prawitz's natural deduction framework, shows completeness for intuitionistic propositional logic in such a system in
\cite{PSH1984b} and for an extension to intuitionistic quantifier logic in \cite{PSH1984a}.
Wansing uses von Kutschera's approach to show the functional completeness of operators for substructural subsystems of intuitionistic propositional logic \cite{Wansing1993a} and of \texttt{N4} \cite[Ch. 7]{Wansing1993b} and also for several normal modal and tense logics with respect to a representation of the generalized rules in display sequent calculi \cite{Wansing1996}.
Braüner \cite{Brauner2005} does the same for hybridized versions of certain modal logics.
It is especially von Kutschera's later paper \cite{Kutschera1969} that is interesting to consider from a \emph{bilateralist} proof-theoretic semantics point of view and also for the present purpose.\footnote{See \cite{TranslationKutschera} for a translation of this paper as well as \cite{Ayhancomment} for a comment on its relevance and its connection to \cite{Kutschera1968}.}
In this paper von Kutschera is concerned with giving a concept of refutation that is -- other than in intuitionistic logic -- not derivative on a concept of proof but on a par with it.
This is a very bilateralist stance, since bilateralism is the view that in proof-theoretic semantics it is not only the rules governing proofs of complex formulas but also the rules governing their refutations that need to be considered as meaning-determining and, importantly, that these concepts should be considered both as primitive, not one reducible to the other.\footnote{More often in bilateralism the terms `assertion' and `denial' are used as the relevant concepts but like von Kutschera we will stick to the terms `proof' and `refutation'.}
Accordingly, in \cite{Kutschera1969}, in order to define arbitrary $n$n-ary connectives, he gives generalized rule schemata not only to introduce \emph{proofs} but also to introduce \emph{refutations}.
We will start with a brief introduction to our system, considering the idea of connexive logics in general, the fact that \texttt{C} is a contradictory logic and the relevant differences to \texttt{N4} and to von Kutschera's approach (Section 2).
To provide a basis for our proof of functional completeness, we will firstly consider generalized derivability schemata for unspecified logical operators as von Kutschera's schemata but with a different background notion of higher-level refutability reflecting the connexive reading of consequence (Section 3-4).
Subsequently, we trace von Kutschera's proof of functional completeness for the \texttt{N4} connectives, yielding the result that the connectives of \texttt{C} are functionally complete for these generalized schemata (Section 5-6).
|
Introduction
| false
|
2507.06854
| 1
|
95,367
|
There has been a recent increase in attention with respect to studying connexive logics, resulting in a number of papers published on the subject.\footnote{See, e.g., the contributions in the special issue of \emph{Studia Logica} on \emph{Frontiers of Connexive Logic} \cite{OmoriWansingSI}.}
The general idea of connexivity that should be implemented by these systems, is that no formula provably implies or is implied by its own negation. Formally this is usually realized by validating so-called \emph{Aristotle's} (AT) and \emph{Boethius' Theses} (BT), while not validating symmetry of implication.
Hence, the following is the most common, agreed-upon definition:
\begin{definition}[Connexive logics]
A logic is connexive iff
1.) the following formulas are provable in the system:
\begin{itemize}
\item AT: ${\sim} ({\sim} A \rightarrow A)$,
\item AT': ${\sim} (A \rightarrow {\sim} A)$,
\item BT: $(A \rightarrow B) \rightarrow {\sim}(A \rightarrow {\sim} B)$,
\item BT': $(A \rightarrow {\sim} B) \rightarrow {\sim}(A \rightarrow B)$, and
\end{itemize}
2.) the system satisfies non-symmetry of implication, i.e., $(A\rightarrow B) \rightarrow (B \rightarrow A)$ is not provable.
\end{definition}\begin{definition}[Connexive logics]
A logic is connexive iff
1.) the following formulas are provable in the system:
\begin{itemize}
\item AT: ${\sim} ({\sim} A \rightarrow A)$,
\item AT': ${\sim} (A \rightarrow {\sim} A)$,
\item BT: $(A \rightarrow B) \rightarrow {\sim}(A \rightarrow {\sim} B)$,
\item BT': $(A \rightarrow {\sim} B) \rightarrow {\sim}(A \rightarrow B)$, and
\end{itemize}
2.) the system satisfies non-symmetry of implication, i.e., $(A\rightarrow B) \rightarrow (B \rightarrow A)$ is not provable.
\end{definition}
A logic is connexive iff
1.) the following formulas are provable in the system:
\item AT: ${\sim} ({\sim} A \rightarrow A)${\sim}\sim ({\sim}\sim A \rightarrow A),
\item AT': ${\sim} (A \rightarrow {\sim} A)${\sim}\sim (A \rightarrow {\sim}\sim A),
\item BT: $(A \rightarrow B) \rightarrow {\sim}(A \rightarrow {\sim} B)$(A \rightarrow B) \rightarrow {\sim}\sim(A \rightarrow {\sim}\sim B),
\item BT': $(A \rightarrow {\sim} B) \rightarrow {\sim}(A \rightarrow B)$(A \rightarrow {\sim}\sim B) \rightarrow {\sim}\sim(A \rightarrow B), and
2.) the system satisfies non-symmetry of implication, i.e., $(A\rightarrow B) \rightarrow (B \rightarrow A)$(A\rightarrow B) \rightarrow (B \rightarrow A) is not provable.
We will concentrate here on the system \texttt{C}, introduced in \cite{Wansing2005}, which is obtained by changing the falsification conditions of implication in Nelson's four-valued constructive logic with strong negation, \texttt{N4} \cite{AN1984, Nelson1949}.\footnote{To give more context and references for the system \texttt{C}: See, e.g., \cite{Wansing2024} and \cite{WansingWeber} for motivations for \texttt{C}, \cite{Olkhovikov2023}, where a direct completeness proof for 1st-order \texttt{C} is given, or \cite{FazioOdintsov2024} for an algebraic investigation of the system. For a sequent calculus for \texttt{C}, see Appendix.}
The following is a natural deduction calculus for the propositional fragment of \texttt{C}, which is adapted from a bilateral version given in \cite[p. 419f.]{Wansing2016}.
Although in the paper introducing this system \cite{Wansing2005}, the aim is to present a connexive \emph{modal} logic, we will not be concerned with modal operators here and rather concentrate on the way this system is constructed from \texttt{N4}.
\vspace{0.5cm}
{\Large\textbf{\texttt{NC}}}\Large\textbf{\texttt{NC}}
\vspace{0.2cm}
\textbf{Rules for ${\sim}$:}
\vspace{0.2cm}
\quad
\infer[\scriptstyle{\sim}\sim {\sim}\sim I]{{\sim} {\sim} A}{\sim}\sim {\sim}\sim A{\infer*{A}{\Gamma}}\infer*{A}A{\Gamma}\Gamma
\quad
\infer[\scriptstyle{\sim}\sim{\sim}\sim E]{A}A{
\infer*{{\sim} {\sim} A}{\Gamma}}
\infer*{{\sim} {\sim} A}{\sim}\sim {\sim}\sim A{\Gamma}\Gamma
\vspace{0.2cm}
\textbf{Rules for $\wedge$:}
\vspace{0.2cm}
\quad
\infer[\scriptstyle\wedge I]{A \wedge B}A \wedge B{\;\infer*{A}{\Gamma} \quad \quad \infer*{B}{\Delta}}\;\infer*{A}A{\Gamma}\Gamma \quad \quad \infer*{B}B{\Delta}\Delta
\quad
\infer[\scriptstyle\wedge E_{1}1]{A}A{\infer*{A \wedge B}{\Gamma}}\infer*{A \wedge B}A \wedge B{\Gamma}\Gamma
\quad
\infer[\scriptstyle\wedge E_{2}2]{B}B{\infer*{A \wedge B}{\Gamma}}\infer*{A \wedge B}A \wedge B{\Gamma}\Gamma
\vspace{0.2cm}
\quad
\infer[\scriptstyle{\sim}\sim\wedge I_{1}1]{{\sim}(A \wedge B)}{\sim}\sim(A \wedge B){\infer*{{\sim} A}{\Gamma}}\infer*{{\sim} A}{\sim}\sim A{\Gamma}\Gamma
\quad \quad
\infer[\scriptstyle{\sim}\sim\wedge I_{2}2]{{\sim}(A \wedge B)}{\sim}\sim(A \wedge B){\infer*{{\sim} B}{\Gamma}}\infer*{{\sim} B}{\sim}\sim B{\Gamma}\Gamma
\quad
\hspace{-0.5cm}
\infer[\scriptstyle{\sim}\sim\wedge E]{C}C{\;\;\;\;\;\;\;\;\;\infer*{{\sim}(A \wedge B)}{\Gamma} \quad \infer*{C}{[{\sim} A], \Delta} \quad \infer*{C}{[{\sim} B], \Theta }}\;\;\;\;\;\;\;\;\;\infer*{{\sim}(A \wedge B)}{\sim}\sim(A \wedge B){\Gamma}\Gamma \quad \infer*{C}C{[{\sim} A], \Delta}[{\sim}\sim A], \Delta \quad \infer*{C}C{[{\sim} B], \Theta }[{\sim}\sim B], \Theta
\vspace{0.2cm}
\textbf{Rules for $\vee$:}
\vspace{0.2cm}
\quad
\infer[\scriptstyle\vee I_{1}1]{A \vee B}A \vee B{\infer*{A}{\Gamma}}\infer*{A}A{\Gamma}\Gamma
\quad \quad
\infer[\scriptstyle\vee I_{2}2]{A \vee B}A \vee B{\infer*{B}{\Gamma}}\infer*{B}B{\Gamma}\Gamma
\quad
\hspace{-0.5cm}
\infer[\scriptstyle\vee E]{C}C{\;\;\;\;\;\;\;\;\;\infer*{A \vee B}{\Gamma} \quad \infer*{C}{ [A], \Delta} \quad \infer*{C}{[B], \Theta}}\;\;\;\;\;\;\;\;\;\infer*{A \vee B}A \vee B{\Gamma}\Gamma \quad \infer*{C}C{ [A], \Delta} [A], \Delta \quad \infer*{C}C{[B], \Theta}[B], \Theta
\vspace{0.2cm}
\quad
\infer[\scriptstyle{\sim}\sim\vee I]{{\sim}(A \vee B)}{\sim}\sim(A \vee B){\;\infer*{{\sim} A}{\Gamma}\quad \quad\infer*{{\sim} B}{\Delta} }\;\infer*{{\sim} A}{\sim}\sim A{\Gamma}\Gamma\quad \quad\infer*{{\sim} B}{\sim}\sim B{\Delta}\Delta
\quad \quad \quad
\infer[\scriptstyle{\sim}\sim\vee E_{1}1]{{\sim} A}{\sim}\sim A{\infer*{{\sim}(A \vee B)}{\Gamma}}\infer*{{\sim}(A \vee B)}{\sim}\sim(A \vee B){\Gamma}\Gamma
\quad \quad
\infer[\scriptstyle{\sim}\sim\vee E_{2}2]{{\sim} B}{\sim}\sim B{\infer*{{\sim}(A \vee B)}{\Gamma}}\infer*{{\sim}(A \vee B)}{\sim}\sim(A \vee B){\Gamma}\Gamma
\vspace{0.2cm}
\textbf{Rules for $\rightarrow$:}
\vspace{0.2cm}
\quad
\infer[\scriptstyle\rightarrow I]{A \rightarrow B}A \rightarrow B{
\infer*{B}{[A],\Gamma}}
\infer*{B}B{[A],\Gamma}[A],\Gamma
\quad \hspace{-0.35cm}
\infer[\scriptstyle\rightarrow E]{B}B{\infer*{A}{\Gamma} \ \quad \infer*{A \rightarrow B}{\Delta}}\infer*{A}A{\Gamma}\Gamma \ \quad \infer*{A \rightarrow B}A \rightarrow B{\Delta}\Delta
\quad \hspace{-0.35cm}
\infer[\scriptstyle{\sim}\sim\rightarrow I]{{\sim}(A \rightarrow B)}{\sim}\sim(A \rightarrow B){
\infer*{{\sim} B}{[A],\Gamma}}
\infer*{{\sim} B}{\sim}\sim B{[A],\Gamma}[A],\Gamma
\quad \hspace{-0.35cm}
\infer[\scriptstyle{\sim}\sim\rightarrow E]{{\sim} B}{\sim}\sim B{\infer*{A}{\Gamma} \ \quad \infer*{{\sim}(A \rightarrow B)}{\Delta}}\infer*{A}A{\Gamma}\Gamma \ \quad \infer*{{\sim}(A \rightarrow B)}{\sim}\sim(A \rightarrow B){\Delta}\Delta
The difference between a natural deduction system for \texttt{N4}, as can be found in \cite[p. 97]{Prawitz1965}, for example, and \texttt{NC} lies in the rules for negated implication.
We obtain a connexive system by ${\sim\rightarrow} I${\sim\rightarrow}\sim\rightarrow I and ${\sim\rightarrow} E${\sim\rightarrow}\sim\rightarrow E, as given above, replacing these rules in \texttt{N4}:
\quad
\infer[\scriptstyle\sim\rightarrow I]{\sim(A \rightarrow B)}\sim(A \rightarrow B){{A}\quad\quad {\sim B}}{A}A\quad\quad {\sim B}\sim B
\quad \quad
\infer[\scriptstyle\sim\rightarrow E_{1}1]{A}A{\sim(A \rightarrow B)}\sim(A \rightarrow B)
\quad \quad
\infer[\scriptstyle\sim\rightarrow E_{2}2]{\sim B}\sim B{\sim(A \rightarrow B)}\sim(A \rightarrow B)
Von Kutschera takes \texttt{N4} to express naturally a concept of direct refutation, represented by the connective ${\sim}${\sim}\sim, which we call `strong negation'.
However, with \texttt{C} being connexive, the interpretation of what it means to refute an implication $A \rightarrow B$A \rightarrow B is very different from \texttt{N4}, where this means to have a proof of $A$A and ${\sim} B${\sim}\sim B.
With a connexive implication refuting an implication $A \rightarrow B$A \rightarrow B means to derive a refutation of $B$B from the assumption $A$A.\footnote{To be more precise, actually this notion captures what is sometimes called a ``hyperconnexive'' understanding of implication, which is not only reflected in the validity of Boethius' thesis $(A\rightarrow B) \rightarrow {\sim}(A\rightarrow {\sim} B)$ as a characteristic principle of connexive logic, but also in the validity of its converse. For remarks and an overview on the differing terminology in connexive logic, see \cite{WansingOmori2024}.}
Thus, this needs to be reflected in a generalized conception of `higher-order' refutability, too.
Another peculiarity arising from this is that \texttt{C} is a \emph{contradictory logic}.
We use this term for logics going beyond paraconsistency by being not only not explosive but also containing contradictory formulas (here: formulas of the form $A$A and ${\sim} A${\sim}\sim A) in their set of theorems while still being non-trivial.\footnote{See, e.g., \cite{Wansingforthcoming}, where it is argued that it is theoretically rational to work with non-trivial negation inconsistent logics and a list of such logics is presented, or \cite{NikiWansing2023}, for examples and a classification of provable contradictions in \texttt{C} and an extension of \texttt{C}.} To give a simple example of such a provable contradiction in \texttt{C}:
\vspace{0.2cm}
\begin{center}
\quad
\infer[\scriptstyle \rightarrow I]{(A \wedge {\sim} A) \rightarrow A}{\infer[\scriptstyle \wedge E_1]{A}{{[A \wedge {\sim} A]}}}
\quad \quad
\infer[\scriptstyle {\sim}\rightarrow I]{\sim((A \wedge {\sim} A) \rightarrow A)}{\infer[\scriptstyle \wedge E_2]{{\sim} A}{{[A \wedge {\sim} A]}}}
\end{center}\begin{center}
\quad
\infer[\scriptstyle \rightarrow I]{(A \wedge {\sim} A) \rightarrow A}{\infer[\scriptstyle \wedge E_1]{A}{{[A \wedge {\sim} A]}}}
\quad \quad
\infer[\scriptstyle {\sim}\rightarrow I]{\sim((A \wedge {\sim} A) \rightarrow A)}{\infer[\scriptstyle \wedge E_2]{{\sim} A}{{[A \wedge {\sim} A]}}}
\end{center}
\quad
\infer[\scriptstyle \rightarrow I]{(A \wedge {\sim} A) \rightarrow A}(A \wedge {\sim}\sim A) \rightarrow A{\infer[\scriptstyle \wedge E_1]{A}{{[A \wedge {\sim} A]}}}\infer[\scriptstyle \wedge E_1]{A}A{{[A \wedge {\sim} A]}}{[A \wedge {\sim} A]}[A \wedge {\sim}\sim A]
\quad \quad
\infer[\scriptstyle {\sim}\sim\rightarrow I]{\sim((A \wedge {\sim} A) \rightarrow A)}\sim((A \wedge {\sim}\sim A) \rightarrow A){\infer[\scriptstyle \wedge E_2]{{\sim} A}{{[A \wedge {\sim} A]}}}\infer[\scriptstyle \wedge E_2]{{\sim} A}{\sim}\sim A{{[A \wedge {\sim} A]}}{[A \wedge {\sim} A]}[A \wedge {\sim}\sim A]
The proof of functional completeness for \texttt{C} can essentially proceed in the fashion of von Kutschera's, so we do not think it is necessary to provide it in full length here but we will restrict the paper to a sketch of the proof while making the differences explicit.
There is one important conceptual point worth mentioning, though.
When discussing under what conditions we can give rules for refutations of arbitrary connectives, von Kutschera says that ``[o]bviously, these rules [for refutation] cannot be devised independently of [the rules for proofs] if the consistency of the Gentzen calculi is to be maintained when these rules are added'' \cite[p. 108, (translated by the authors)]{Kutschera1969}.
What he fails to mention in this context is that our conception of how precisely to refute inferences plays a crucial role here.
We also give the refutation rules dependently on the proof rules in the same manner as it is done in \cite{Kutschera1969} and we do get the completeness result for the \emph{inconsistent} logic \texttt{C}.
So, in light of our adjustments of von Kutschera's approach for \texttt{C} (Section 4), the dependence of the refutation rules on the proof rules seems to be an independent matter, separate from any considerations about consistency.
It is rather the underlying notion of refutation that is important here and with respect to \emph{that} our approaches are obviously essentially different.
While von Kutschera (without any further explanation or justification) assumes a notion of refutation that captures the, in many logics, usual way of interpreting a negated implication, i.e., the refutability of a derivation going from $A$A to $B$B is defined by the provability of $A$A and the refutability of $B$B, our notion will capture the connexive understanding of implication.
Thus, the refutability of a derivation going from $A$A to $B$B will be defined by the refutability of $B$B given the assumed provability of $A$A.
So, it is rather von Kutschera's underlying assumption that derivability should basically behave as implication does in \texttt{N4} that is critical for consistency, not the factor of dependent definitions between proofs and refutations.
|
The motivation: A contradictory logic that is functionally complete
| false
|
2507.06854
| 2
|
3
|
\label{TDA1D}
In this section, we sketch the basic ideas of the rapidly expanding field of topological data analysis by considering a simple application in one dimension that simultaneously advances the state of the art in the fundamental area of nonparametric statistical estimation and avoids much of the technical baggage of persistent homology.
\emph{Topological data analysis}\index{topological data analysis}topological data analysis (TDA) has had a profound effect on data science and statistics over the last 15 years. Perhaps the most widely recognized and utilized tool in TDA is \emph{persistent homology}\index{persistent homology}persistent homology \cite{zomorodian2005topology,Ghrist_2008,Carlsson2009,Edelsbrunner2010,GhristEAT,Oudot2015}. The basic idea (Figure \ref{fig:TDA}) is to associate an inclusion-oriented family (i.e., a \emph{filtration}\index{filtration}filtration) of simplicial complexes to a point set in a metric space. Each simplicial complex in the filtration is formed by considering the intersections of balls of a fixed radius about each data point. As the radius varies, different simplicial complexes are produced, and their homologies are computed.
\begin{figure}[htb]
\includegraphics[trim = 15mm 43mm 15mm 27mm, clip, keepaspectratio, width=\textwidth]{circles.pdf}
%%\centerline{\epsfig{figure=cat.eps,width=.5\textheight,height=.4\textwidth}}
\caption[Topological persistence]{The topology of a data set can be probed at different scales. Here, we consider a sample of 100 uniformly distributed points in a thin annulus about the unit circle. From left to right, we place disks of radius $0.1$, $0.15$, and $0.95$ around each point. The topology of the data set is morally that of a circle, and the (persistent) homology of simplicial complexes formed from the intersections of disks reveals this: a 1-homology class ``persists'' over an interval slightly bigger than $[.15, .95]$.}
\label{fig:TDA}
\end{figure}
\includegraphics[trim = 15mm 43mm 15mm 27mm, clip, keepaspectratio, width=\textwidth]{circles.pdf}
\caption[Topological persistence]{The topology of a data set can be probed at different scales. Here, we consider a sample of 100 uniformly distributed points in a thin annulus about the unit circle. From left to right, we place disks of radius $0.1$, $0.15$, and $0.95$ around each point. The topology of the data set is morally that of a circle, and the (persistent) homology of simplicial complexes formed from the intersections of disks reveals this: a 1-homology class ``persists'' over an interval slightly bigger than $[.15, .95]$.}The topology of a data set can be probed at different scales. Here, we consider a sample of 100 uniformly distributed points in a thin annulus about the unit circle. From left to right, we place disks of radius $0.1$0.1, $0.15$0.15, and $0.95$0.95 around each point. The topology of the data set is morally that of a circle, and the (persistent) homology of simplicial complexes formed from the intersections of disks reveals this: a 1-homology class ``persists'' over an interval slightly bigger than $[.15, .95]$[.15, .95].
\label{fig:TDA}
Although the theory of topological persistence involves a considerable amount of algebra for bookkeeping associated to the ``births'' and ``deaths'' of homology classes as a function of the radius/filtration parameter, in practice simply treating the Betti numbers as functions of that parameter gives considerable information. Along similar lines, we can consider how other topological invariants behave as a function of scale.
\begin{figure}[htb]
\begin{center}
\includegraphics[trim = 40mm 109mm 40mm 100mm, clip, keepaspectratio, width=.48\textwidth]{tmeArea.pdf}
\includegraphics[trim = 40mm 109mm 40mm 100mm, clip, keepaspectratio, width=.48\textwidth]{tmeLine.pdf}
%%\centerline{\epsfig{figure=cat.eps,width=.5\textheight,height=.4\textwidth}}
\end{center}
\caption[Topological mixture estimation]{Topological mixture estimation. Left panels: area plots of (top) initial and (bottom) information-theoretically optimized unimodal decompositions of an estimated probability distribution. Right panels: line plots of the same decompositions. The bandwidth for the kernel density estimate \footnote{A kernel density estimate is a sort of ``smooth histogram'' that represents sample data by an average of copies of a ``kernel'' probability distribution centered around the data points. The bandwidth of a kernel density estimate is a scaling factor for the kernel: small bandwidths scale the kernel to be narrow and tall, and large bandwidths scale the kernel to be wide and short.} for the distribution and the number of unimodal mixture components are both determined using the same topological considerations.}
\label{fig:TME}
\end{figure}
\includegraphics[trim = 40mm 109mm 40mm 100mm, clip, keepaspectratio, width=.48\textwidth]{tmeArea.pdf}
\includegraphics[trim = 40mm 109mm 40mm 100mm, clip, keepaspectratio, width=.48\textwidth]{tmeLine.pdf}
\caption[Topological mixture estimation]{Topological mixture estimation. Left panels: area plots of (top) initial and (bottom) information-theoretically optimized unimodal decompositions of an estimated probability distribution. Right panels: line plots of the same decompositions. The bandwidth for the kernel density estimate \footnote{A kernel density estimate is a sort of ``smooth histogram'' that represents sample data by an average of copies of a ``kernel'' probability distribution centered around the data points. The bandwidth of a kernel density estimate is a scaling factor for the kernel: small bandwidths scale the kernel to be narrow and tall, and large bandwidths scale the kernel to be wide and short.} for the distribution and the number of unimodal mixture components are both determined using the same topological considerations.}Topological mixture estimation. Left panels: area plots of (top) initial and (bottom) information-theoretically optimized unimodal decompositions of an estimated probability distribution. Right panels: line plots of the same decompositions. The bandwidth for the kernel density estimate \footnote{A kernel density estimate is a sort of ``smooth histogram'' that represents sample data by an average of copies of a ``kernel'' probability distribution centered around the data points. The bandwidth of a kernel density estimate is a scaling factor for the kernel: small bandwidths scale the kernel to be narrow and tall, and large bandwidths scale the kernel to be wide and short.} for the distribution and the number of unimodal mixture components are both determined using the same topological considerations.
\label{fig:TME}
Call $\phi : \mathbb{R}^n \rightarrow [0,\infty)$\phi : \mathbb{R}^n \rightarrow [0,\infty) \emph{unimodal}\index{unimodal}unimodal if $\phi$\phi is continuous and the excursion set $\phi^{-1}([y,\infty))$\phi^{-1}-1([y,\infty)) is contractible (i.e., homotopy equivalent to a point) for all $0 < y \le \max \phi$0 < y \le \max \phi. For $n = 1$n = 1, contractibility means that these excursion sets are all intervals, which coincides with the intuitive notion of unimodality. For $f : \mathbb{R}^n \rightarrow [0,\infty)$f : \mathbb{R}^n \rightarrow [0,\infty) sufficiently nice, define the \emph{unimodal category}\index{unimodal!category}unimodal!category of $f$f to be the smallest number $M$M of functions such that $f$f admits a \emph{unimodal decomposition}\index{unimodal!decomposition}unimodal!decomposition of the form $f = \sum_{m=1}^M \pi_m \phi_m$f = \sum_{m=1}m=1^M \pi_m \phi_m for some $\pi > 0$\pi > 0, $\sum_m \pi_m = 1$\sum_m \pi_m = 1, and $\phi_m$\phi_m unimodal \cite{GhristEAT}.
The unimodal category is a topological (homeomorphism) invariant and a ``sweep'' algorithm due to Baryshnikov and Ghrist efficiently produces a unimodal decomposition in $n = 1$n = 1. \footnote{The case $n = 2$ is still beyond the reach of current techniques, and only partial results are known. Moreover, for $n$ sufficiently large, there is provably no algorithm for computing the unimodal category!} As Figure \ref{fig:TME} demonstrates, the unimodal category can be much less than the number of extrema.
The unimodal category of a kernel density estimate for a probability distribution can be used to select an appropriate bandwidth for sample data and, as shown in Figure \ref{fig:TME}, to decompose the resulting estimated distribution into well-behaved unimodal components \emph{using no externally supplied parameters whatsoever} \cite{Huntsman2018}. The key ideas behind \emph{topological mixture estimation}\index{topological mixture estimation}topological mixture estimation are to identify the most common unimodal category as a function of bandwidth and to exploit convexity properties of the mutual information between the mixture weights and the distribution itself. The result is an extremely general (though also computationally expensive) unsupervised learning technique in one dimension that can, e.g. automatically set thresholds for anomaly detectors or determine the number of clusters in data (by taking random projections).
|
Topological data analysis and unsupervised learning in one dimension
| false
|
2008.03299
| 3
|
5
|
\label{Conclusion}
We have only scratched the surface of topological techniques that can be fruitfully applied to problems in the cyber domain. Discrete Morse theory \cite{Scoville_2019}, the algebraic topology of finite topological spaces \cite{barmak2011algebraic}, and connections between simplicial complexes and partially ordered sets \cite{wachs2007poset} provide just a few opportunities for applications that we have not discussed at all here. For example, a notion of a weighted Dowker complex and an associated partial order can be used for \emph{topological differential testing}\index{topological differential testing}topological differential testing to discover files that similar programs handle inconsistently \cite{ambrose2020topological}.
More generally, both discrete and continuous topological methods can provide unique capabilities for problems in the cyber domain. The analysis of concurrent protocols and programs highlights this: while simplicial complexes have been used to solve problems in concurrency \cite{herlihy2014distributed}, the entire (recently developed) theory of directed topology traces its origin to static analysis of concurrent programs \cite{fajstrup2016directed}.
In short, while there are many cyber-oriented problems that present a large attack surface for mainstream topological data analysis, the space of applicable techniques is much larger. Cyber problems are likely to continue to motivate future developments in topology, both theoretical and applied.
\subsection*{Acknowledgement}
The authors thank Samir Chowdhury, Fabrizio Romano Genovese, Jelle Herold, and Matvey Yutin for helpful discussions; Greg Sadosuk for producing Figure \ref{fig:intel_cfg} and its attendant code, and Richard Latimer for providing code and data relating to sorting networks.
This research was partially supported with funding from the Defense Advanced Research Projects Agency (DARPA) via Federal contracts HR001115C0050 and HR001119C0072. The views, opinions, and/or findings expressed are those of the authors and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government.
\begin{thebibliography}{10}
\bibitem{nsnam}
The {NS-2} network simulator.
\newblock {\tt http://www.nsnam.org/}.
\newblock Accessed: 2016-05-23.
\bibitem{Alves2016}
H.~Alves, B.~Fonseca, and N.~Antunes.
\newblock Software metrics and security vulnerabilities: dataset and
exploratory study.
\newblock In {\em IDCC}, 2016.
\bibitem{ambrose2020topological}
K.~Ambrose, S.~Huntsman, M.~Robinson, and M.~Yutin.
\newblock Topological differential testing.
\newblock {\em arXiv preprint arXiv:2003.00976}, 2020.
\bibitem{Arulselvan_2009}
A.~Arulselvan, C.~W. Commander, L.~Elefteriadou, and P.~M. Pardalos.
\newblock Detecting critical nodes in sparse graphs.
\newblock {\em Comp. Operations Res.}, 36(7):2193--2200, 2009.
\bibitem{Atkin1974}
R.~Atkin.
\newblock {\em Mathematical Structure in Human Affairs}.
\newblock Heinemann, 1974.
\bibitem{ballard2019geometry}
G.~Ballard, C.~Ikenmeyer, J.M. Landsberg, and N.~Ryder.
\newblock The geometry of rank decompositions of matrix multiplication {II}: 3
$\times$ 3 matrices.
\newblock {\em J. Pure Appl. Algebra}, 223(8):3205--3224, 2019.
\bibitem{barmak2011algebraic}
J.~A. Barmak.
\newblock {\em Algebraic Topology of Finite Topological Spaces and
Applications}.
\newblock Springer, 2011.
\bibitem{berger2019equivalent}
G.~O. Berger, , P.-A. Absil, L.~{De Lathauwer}, R.~M. Jungers, and M.~{Van
Barel}.
\newblock Equivalent polyadic decompositions of matrix multiplication tensors.
\newblock {\em arXiv preprint arXiv:1902.03950}, 2019.
\bibitem{Carlsson2009}
G.~Carlsson.
\newblock Topology and data.
\newblock {\em Bull. Amer. Math. Soc.}, 46:255--308, 2009.
\bibitem{Chiang_2007}
M.~Chiang, S.~Low, R.~Calderbank, and J.~Doyle.
\newblock Layering as optimization decomposition: a mathematical theory of
network architectures.
\newblock {\em Proc. IEEE}, 95(1), January 2007.
\bibitem{chokaev2018two}
B.~V. Chokaev and G.~N. Shumkin.
\newblock Two bilinear (3 $\times$ 3)-matrix multiplication algorithms of
complexity 25.
\newblock {\em Moscow U. Comp. Math. Cyber.}, 42(1):23--30, 2018.
\bibitem{Chowdhury2019}
S.~Chowdhury, T.~Gebhart, S.~Huntsman, and M.~Yutin.
\newblock Path homologies of deep feedforward networks.
\newblock In {\em ICMLA}, 2019.
\bibitem{Chowdhury2018}
S.~Chowdhury and F.~M\'emoli.
\newblock A functorial {D}owker theorem and persistent homology of asymmetric
networks.
\newblock {\em J. Appl. Comp. Topology}, 2(1):115, 2018.
\bibitem{courtois2011new}
N.~T. Courtois, G.~V. Bard, and D.~Hulme.
\newblock A new general-purpose method to multiply 3 $\times$ 3 matrices using
only 23 multiplications.
\newblock {\em arXiv preprint arXiv:1108.2830}, 2011.
\bibitem{DiSumma_2011}
M.~{Di Summa}, A.~Grosso, and M.~Locatelli.
\newblock Complexity of the critical node problem over trees.
\newblock {\em Comp. Operations Res.}, 38(12):1766--1774, 2011.
\bibitem{dinh2012}
T.~N. Dinh, Y.~Xuan, M.~T. Thai, P.~M. Pardalos, and T.~Znati.
\newblock On new approaches of assessing network vulnerability: hardness and
approximation.
\newblock {\em IEEE/ACM Trans. Networking}, 20(2):609--619, 2012.
\bibitem{Dowker1952}
C.~H. Dowker.
\newblock Homology groups of relations.
\newblock {\em Ann. Math.}, 56:84, 1952.
\bibitem{Du2019}
X.~Du, B.~Chen, Y.~Li, J.~Guo, Y.~Zhou, Y.~Liu, and Y.~Jiang.
\newblock {LEOPARD}: identifying vulnerable code for vulnerability assessment
through program metrics.
\newblock In {\em ICSE}, 2019.
\bibitem{Dullien2009}
T.~Dullien and S.~Porst.
\newblock {REIL}: a platform-independent intermediate representation of
disassembled code for static code analysis.
\newblock In {\em CanSecWest}, 2009.
\bibitem{Duran2011}
D.~Duran, D.~Weston, and M.~Miller.
\newblock Targeted taint driven fuzzing using software metrics.
\newblock In {\em CanSecWest}, 2011.
\bibitem{Eagle2011}
C.~Eagle.
\newblock {\em The IDA Pro Book: The Unofficial Guide to the World's Most
Popular Disassembler}.
\newblock No Starch Press, 2011.
\bibitem{Ebert2016}
C.~Ebert and J.~Cain.
\newblock Cyclomatic complexity.
\newblock {\em IEEE Soft.}, 33:27, 2016.
\bibitem{Edelsbrunner2010}
H.~Edelsbrunner and J.~L. Harer.
\newblock {\em Computational Topology: An Introduction}.
\newblock AMS, 2010.
\bibitem{Erdmann2017}
M.~Erdmann.
\newblock Topology of privacy: lattice structures and information bubbles for
inference and obfuscation.
\newblock {\em arXiv preprint arXiv:1712.04130}, 2017.
\bibitem{fajstrup2016directed}
L.~Fajstrup, E.~Goubault, E.~Haucourt, S.~Mimram, and M.~Raussen.
\newblock {\em Directed Algebraic Topology and Concurrency}.
\newblock Springer, 2016.
\bibitem{Ghrist_2008}
R.~Ghrist.
\newblock Barcodes: the persistent topology of data.
\newblock {\em Bull. Amer. Math. Soc.}, 45(1):61, 2008.
\bibitem{GhristEAT}
R.~Ghrist.
\newblock {\em Elementary Applied Topology}.
\newblock Createspace, 2014.
\bibitem{Ghrist2012}
R.~Ghrist, D.~Lipsky, J.~Derenick, and A.~Speranzon.
\newblock Topological landmark-based navigation and mapping.
\newblock {\em preprint}, 2012.
\bibitem{Grigoryan2018}
A.~Grigor'yan, R.~Jimenez, Yu. Muranov, and S.-T. Yau.
\newblock On the path homology theory of digraphs and {Eilenberg-Steenrod}
axioms.
\newblock {\em Homology Homotopy Appl.}, 20:179, 2018.
\bibitem{Grigoryan2018b}
A.~Grigor'yan, Yu. Muranov, V.~Vershinin, and S.-T. Yau.
\newblock Path homology theory of multigraphs and quivers.
\newblock {\em Forum Math.}, 30:1319, 2018.
\bibitem{Grigoryan2014}
A.~Grigor'yan, Yu. Muranov, and S.-T. Yau.
\newblock Graphs associated with simplicial complexes.
\newblock {\em Homology Homotopy Appl.}, 16:295, 2014.
\bibitem{Grigoryan2017}
A.~Grigor'yan, Yu. Muranov, and S.-T. Yau.
\newblock Homologies of graphs and {K}\"unneth formulas.
\newblock {\em Comm. Anal. Geom.}, 25:969, 2017.
\bibitem{Grigoryan2012}
A.~Grigor'yan, L.~Yong, Yu. Muranov, and S.-T. Yau.
\newblock Homologies of path complexes and digraphs.
\newblock {\em arXiv preprint arXiv:1207.2834}, 2012.
\bibitem{Grigoryan2014b}
A.~Grigor'yan, L.~Yong, Yu. Muranov, and S.-T. Yau.
\newblock Homotopy theory for digraphs.
\newblock {\em Pure Appl. Math. Quart.}, 10:619, 2014.
\bibitem{Grigoryan2015}
A.~Grigor'yan, L.~Yong, Yu. Muranov, and S.-T. Yau.
\newblock Cohomology of digraphs and (undirected) graphs.
\newblock {\em Asian J. Math.}, 19:887, 2015.
\bibitem{Gueye_2010}
A.~Gueye, J.~C. Walrand, and V.~Anantharam.
\newblock Design of network topology in an adversarial environment.
\newblock In {\em Decision and Game Theory for Security}, pages 1--20.
Springer, 2010.
\bibitem{Hatcher_2002}
A.~Hatcher.
\newblock {\em Algebraic Topology}.
\newblock Cambridge, 2002.
\bibitem{herlihy2014distributed}
M.~Herlihy, D.~Kozlov, and S.~Rajsbaum.
\newblock {\em Distributed Computing Through Combinatorial Topology}.
\newblock Morgan Kaufmann, 2014.
\bibitem{heule2019new}
M.~J.~H. Heule, M.~Kauers, and M.~Seidl.
\newblock New ways to multiply 3 $\times$ 3-matrices.
\newblock {\em arXiv preprint arXiv:1905.10192}, 2019.
\bibitem{Huntsman2018}
S.~Huntsman.
\newblock Topological mixture estimation.
\newblock In {\em ICML}, 2018.
\bibitem{Huntsman2020}
S.~Huntsman.
\newblock Generalizing cyclomatic complexity via path homology.
\newblock {\em arXiv preprint arXiv:2003.00944}, 2020.
\bibitem{Iozzo2010}
V.~Iozzo.
\newblock 0-knowledge fuzzing.
\newblock In {\em Black Hat DC}, 2010.
\bibitem{Jain_2003}
K.~Jain, J.~Padhye, V.~Padmanabhan, and L.~Qiu.
\newblock Impact of interference on multi-hop wireless network performance.
\newblock In {\em MobiCom}, 2003.
\bibitem{johnson1986noncommutative}
R.~W. Johnson and A.~M. McLoughlin.
\newblock Noncommutative bilinear algorithms for 3 $\times$ 3 matrix
multiplication.
\newblock {\em SIAM J. Comp.}, 15(2):595--603, 1986.
\bibitem{Joslyn_2016}
C.~Joslyn, B.~Praggastis, E.~Purvine, A.~Sathanur, M.~Robinson, and
S.~Ranshous.
\newblock Local homology dimension as a network science measure.
\newblock In {\em SIAM Workshop on Network Science}, Boston, July 2016.
\bibitem{knuth1997art}
D.~E. Knuth.
\newblock {\em The Art of Computer Programming}, volume~3.
\newblock Pearson, 1997.
\bibitem{laderman1976noncommutative}
J.~D. Laderman.
\newblock A noncommutative algorithm for multiplying 3 $\times$ 3 matrices
using 23 multiplications.
\newblock {\em Bull. Amer. Math. Soc.}, 82(1):126--128, 1976.
\bibitem{landsberg2017geometry}
J.~M. Landsberg.
\newblock {\em Geometry and Complexity Theory}.
\newblock Cambridge, 2017.
\bibitem{Lee_2007}
J.-W. Lee, M.~Chiang, and R.~Calderbank.
\newblock Utility-optimal random-access control.
\newblock {\em IEEE Trans. Wireless Comm.}, 6(7):2741--2751, 2007.
\bibitem{makarov1986algorithm}
Oleg~M Makarov.
\newblock An algorithm for multiplying 3$\times$ 3 matrices.
\newblock {\em USSR Comp. Math. Math. Phys.}, 26(1):179--180, 1986.
\bibitem{McCabe1976}
T.~J. McCabe.
\newblock A complexity measure.
\newblock {\em IEEE Trans. Soft. Eng.}, SE-2:308, 1976.
\bibitem{Medeiros2017}
N.~Medeiros, N.~Ivaki, P.~Costa, and M.~Vieira.
\newblock Software metrics as indicators of security vulnerabilities.
\newblock In {\em ISSRE}, 2017.
\bibitem{Mesnard2016}
F.~Mesnard, \'E. Payet, and W.~Vanhoof.
\newblock Towards a framework for algorithm recognition in binary code.
\newblock In {\em PPDP}, 2016.
\bibitem{Nandagopal_2000}
T.~Nandagopal, T.-E. Kim, X.~Gao, and V.~Bharghavan.
\newblock Achieving {MAC} layer fairness in wireless packet networks.
\newblock In {\em MobiCom}, pages 87--–98, 2002.
\bibitem{nielson1992semantics}
H.~R. Nielson and F.~Nielson.
\newblock {\em Semantics with Applications}.
\newblock Springer, 1992.
\bibitem{Noubir_2004}
G.~Noubir.
\newblock On connectivity in \emph{ad hoc} networks under jamming using
directional antennas and mobility.
\newblock In {\em Wired/Wireless Internet Comm.}, pages 186--200. Springer,
2004.
\bibitem{Oudot2015}
S.~Y. Oudot.
\newblock {\em Persistence Theory: From Quiver Representations to Data
Analysis}.
\newblock AMS, 2015.
\bibitem{richardson1969some}
D.~Richardson.
\newblock Some undecidable problems involving elementary functions of a real
variable.
\newblock {\em J. Symb. Logic}, 33(4):514--520, 1969.
\bibitem{RobinsonGlobalSIP2014}
M.~Robinson.
\newblock Analyzing wireless communication network vulnerability with
homological invariants.
\newblock In {\em GlobalSIP}, 2014.
\bibitem{Robinson_2014}
M.~Robinson.
\newblock {\em Topological Signal Processing}.
\newblock Springer, 2014.
\bibitem{robinson2017sheaf}
M.~Robinson.
\newblock Sheaf and duality methods for analyzing multi-model systems.
\newblock In {\em Recent Applications of Harmonic Analysis to Function Spaces,
Differential Equations, and Data Science}, pages 653--703. Springer, 2017.
\bibitem{Scoville_2019}
N.~A. Scoville.
\newblock {\em Discrete Morse Theory}.
\newblock American Mathematical Society.
\bibitem{Shalaby2017}
M.~Shalaby, T.~Mehrez, A.~El-Mougy, K.~Abdulnasser, and A.~Al-Safty.
\newblock Automatic algorithm recognition of source-code using machine
learning.
\newblock In {\em ICMLA}, 2017.
\bibitem{strassen1969gaussian}
V.~Strassen.
\newblock Gaussian elimination is not optimal.
\newblock {\em Numerische Mathematik}, 13(4):354--356, 1969.
\bibitem{Taherkhani2010}
A.~Taherkhani, A.~Korhonen, and L.~Malmi.
\newblock Recognizing algorithms using language constructs, software metrics,
and roles of variables: an experiment with sorting algorithms.
\newblock {\em Computer J.}, 54:1049--1066, 2010.
\bibitem{wachs2007poset}
M.~L. Wachs.
\newblock Poset topology: tools and applications.
\newblock In E.~Miller, V.~Reiner, and B.~Sturmfels, editors, {\em Geometric
Combinatorics}.
\bibitem{winograd1971multiplication}
S.~Winograd.
\newblock On multiplication of 2 $\times$ 2 matrices.
\newblock {\em Linear Algebra Appl.}, 4(4):381--388, 1971.
\bibitem{Yang_2002}
X.~Yang and N.~Vaidya.
\newblock Priority scheduling in wireless \emph{ad hoc} networks.
\newblock In {\em MobiHoc}, 2002.
\bibitem{yanofsky2017galois}
N.~S. Yanofsky.
\newblock Galois theory of algorithms.
\newblock In {\em Rohit Parikh on Logic, Language and Society}, pages 323--347.
Springer, 2017.
\bibitem{zomorodian2005topology}
A.~J. Zomorodian.
\newblock {\em Topology for Computing}.
\newblock Cambridge, 2005.
\end{thebibliography}\begin{thebibliography}{10}
\bibitem{nsnam}
The {NS-2} network simulator.
\newblock {\tt http://www.nsnam.org/}.
\newblock Accessed: 2016-05-23.
\bibitem{Alves2016}
H.~Alves, B.~Fonseca, and N.~Antunes.
\newblock Software metrics and security vulnerabilities: dataset and
exploratory study.
\newblock In {\em IDCC}, 2016.
\bibitem{ambrose2020topological}
K.~Ambrose, S.~Huntsman, M.~Robinson, and M.~Yutin.
\newblock Topological differential testing.
\newblock {\em arXiv preprint arXiv:2003.00976}, 2020.
\bibitem{Arulselvan_2009}
A.~Arulselvan, C.~W. Commander, L.~Elefteriadou, and P.~M. Pardalos.
\newblock Detecting critical nodes in sparse graphs.
\newblock {\em Comp. Operations Res.}, 36(7):2193--2200, 2009.
\bibitem{Atkin1974}
R.~Atkin.
\newblock {\em Mathematical Structure in Human Affairs}.
\newblock Heinemann, 1974.
\bibitem{ballard2019geometry}
G.~Ballard, C.~Ikenmeyer, J.M. Landsberg, and N.~Ryder.
\newblock The geometry of rank decompositions of matrix multiplication {II}: 3
$\times$ 3 matrices.
\newblock {\em J. Pure Appl. Algebra}, 223(8):3205--3224, 2019.
\bibitem{barmak2011algebraic}
J.~A. Barmak.
\newblock {\em Algebraic Topology of Finite Topological Spaces and
Applications}.
\newblock Springer, 2011.
\bibitem{berger2019equivalent}
G.~O. Berger, , P.-A. Absil, L.~{De Lathauwer}, R.~M. Jungers, and M.~{Van
Barel}.
\newblock Equivalent polyadic decompositions of matrix multiplication tensors.
\newblock {\em arXiv preprint arXiv:1902.03950}, 2019.
\bibitem{Carlsson2009}
G.~Carlsson.
\newblock Topology and data.
\newblock {\em Bull. Amer. Math. Soc.}, 46:255--308, 2009.
\bibitem{Chiang_2007}
M.~Chiang, S.~Low, R.~Calderbank, and J.~Doyle.
\newblock Layering as optimization decomposition: a mathematical theory of
network architectures.
\newblock {\em Proc. IEEE}, 95(1), January 2007.
\bibitem{chokaev2018two}
B.~V. Chokaev and G.~N. Shumkin.
\newblock Two bilinear (3 $\times$ 3)-matrix multiplication algorithms of
complexity 25.
\newblock {\em Moscow U. Comp. Math. Cyber.}, 42(1):23--30, 2018.
\bibitem{Chowdhury2019}
S.~Chowdhury, T.~Gebhart, S.~Huntsman, and M.~Yutin.
\newblock Path homologies of deep feedforward networks.
\newblock In {\em ICMLA}, 2019.
\bibitem{Chowdhury2018}
S.~Chowdhury and F.~M\'emoli.
\newblock A functorial {D}owker theorem and persistent homology of asymmetric
networks.
\newblock {\em J. Appl. Comp. Topology}, 2(1):115, 2018.
\bibitem{courtois2011new}
N.~T. Courtois, G.~V. Bard, and D.~Hulme.
\newblock A new general-purpose method to multiply 3 $\times$ 3 matrices using
only 23 multiplications.
\newblock {\em arXiv preprint arXiv:1108.2830}, 2011.
\bibitem{DiSumma_2011}
M.~{Di Summa}, A.~Grosso, and M.~Locatelli.
\newblock Complexity of the critical node problem over trees.
\newblock {\em Comp. Operations Res.}, 38(12):1766--1774, 2011.
\bibitem{dinh2012}
T.~N. Dinh, Y.~Xuan, M.~T. Thai, P.~M. Pardalos, and T.~Znati.
\newblock On new approaches of assessing network vulnerability: hardness and
approximation.
\newblock {\em IEEE/ACM Trans. Networking}, 20(2):609--619, 2012.
\bibitem{Dowker1952}
C.~H. Dowker.
\newblock Homology groups of relations.
\newblock {\em Ann. Math.}, 56:84, 1952.
\bibitem{Du2019}
X.~Du, B.~Chen, Y.~Li, J.~Guo, Y.~Zhou, Y.~Liu, and Y.~Jiang.
\newblock {LEOPARD}: identifying vulnerable code for vulnerability assessment
through program metrics.
\newblock In {\em ICSE}, 2019.
\bibitem{Dullien2009}
T.~Dullien and S.~Porst.
\newblock {REIL}: a platform-independent intermediate representation of
disassembled code for static code analysis.
\newblock In {\em CanSecWest}, 2009.
\bibitem{Duran2011}
D.~Duran, D.~Weston, and M.~Miller.
\newblock Targeted taint driven fuzzing using software metrics.
\newblock In {\em CanSecWest}, 2011.
\bibitem{Eagle2011}
C.~Eagle.
\newblock {\em The IDA Pro Book: The Unofficial Guide to the World's Most
Popular Disassembler}.
\newblock No Starch Press, 2011.
\bibitem{Ebert2016}
C.~Ebert and J.~Cain.
\newblock Cyclomatic complexity.
\newblock {\em IEEE Soft.}, 33:27, 2016.
\bibitem{Edelsbrunner2010}
H.~Edelsbrunner and J.~L. Harer.
\newblock {\em Computational Topology: An Introduction}.
\newblock AMS, 2010.
\bibitem{Erdmann2017}
M.~Erdmann.
\newblock Topology of privacy: lattice structures and information bubbles for
inference and obfuscation.
\newblock {\em arXiv preprint arXiv:1712.04130}, 2017.
\bibitem{fajstrup2016directed}
L.~Fajstrup, E.~Goubault, E.~Haucourt, S.~Mimram, and M.~Raussen.
\newblock {\em Directed Algebraic Topology and Concurrency}.
\newblock Springer, 2016.
\bibitem{Ghrist_2008}
R.~Ghrist.
\newblock Barcodes: the persistent topology of data.
\newblock {\em Bull. Amer. Math. Soc.}, 45(1):61, 2008.
\bibitem{GhristEAT}
R.~Ghrist.
\newblock {\em Elementary Applied Topology}.
\newblock Createspace, 2014.
\bibitem{Ghrist2012}
R.~Ghrist, D.~Lipsky, J.~Derenick, and A.~Speranzon.
\newblock Topological landmark-based navigation and mapping.
\newblock {\em preprint}, 2012.
\bibitem{Grigoryan2018}
A.~Grigor'yan, R.~Jimenez, Yu. Muranov, and S.-T. Yau.
\newblock On the path homology theory of digraphs and {Eilenberg-Steenrod}
axioms.
\newblock {\em Homology Homotopy Appl.}, 20:179, 2018.
\bibitem{Grigoryan2018b}
A.~Grigor'yan, Yu. Muranov, V.~Vershinin, and S.-T. Yau.
\newblock Path homology theory of multigraphs and quivers.
\newblock {\em Forum Math.}, 30:1319, 2018.
\bibitem{Grigoryan2014}
A.~Grigor'yan, Yu. Muranov, and S.-T. Yau.
\newblock Graphs associated with simplicial complexes.
\newblock {\em Homology Homotopy Appl.}, 16:295, 2014.
\bibitem{Grigoryan2017}
A.~Grigor'yan, Yu. Muranov, and S.-T. Yau.
\newblock Homologies of graphs and {K}\"unneth formulas.
\newblock {\em Comm. Anal. Geom.}, 25:969, 2017.
\bibitem{Grigoryan2012}
A.~Grigor'yan, L.~Yong, Yu. Muranov, and S.-T. Yau.
\newblock Homologies of path complexes and digraphs.
\newblock {\em arXiv preprint arXiv:1207.2834}, 2012.
\bibitem{Grigoryan2014b}
A.~Grigor'yan, L.~Yong, Yu. Muranov, and S.-T. Yau.
\newblock Homotopy theory for digraphs.
\newblock {\em Pure Appl. Math. Quart.}, 10:619, 2014.
\bibitem{Grigoryan2015}
A.~Grigor'yan, L.~Yong, Yu. Muranov, and S.-T. Yau.
\newblock Cohomology of digraphs and (undirected) graphs.
\newblock {\em Asian J. Math.}, 19:887, 2015.
\bibitem{Gueye_2010}
A.~Gueye, J.~C. Walrand, and V.~Anantharam.
\newblock Design of network topology in an adversarial environment.
\newblock In {\em Decision and Game Theory for Security}, pages 1--20.
Springer, 2010.
\bibitem{Hatcher_2002}
A.~Hatcher.
\newblock {\em Algebraic Topology}.
\newblock Cambridge, 2002.
\bibitem{herlihy2014distributed}
M.~Herlihy, D.~Kozlov, and S.~Rajsbaum.
\newblock {\em Distributed Computing Through Combinatorial Topology}.
\newblock Morgan Kaufmann, 2014.
\bibitem{heule2019new}
M.~J.~H. Heule, M.~Kauers, and M.~Seidl.
\newblock New ways to multiply 3 $\times$ 3-matrices.
\newblock {\em arXiv preprint arXiv:1905.10192}, 2019.
\bibitem{Huntsman2018}
S.~Huntsman.
\newblock Topological mixture estimation.
\newblock In {\em ICML}, 2018.
\bibitem{Huntsman2020}
S.~Huntsman.
\newblock Generalizing cyclomatic complexity via path homology.
\newblock {\em arXiv preprint arXiv:2003.00944}, 2020.
\bibitem{Iozzo2010}
V.~Iozzo.
\newblock 0-knowledge fuzzing.
\newblock In {\em Black Hat DC}, 2010.
\bibitem{Jain_2003}
K.~Jain, J.~Padhye, V.~Padmanabhan, and L.~Qiu.
\newblock Impact of interference on multi-hop wireless network performance.
\newblock In {\em MobiCom}, 2003.
\bibitem{johnson1986noncommutative}
R.~W. Johnson and A.~M. McLoughlin.
\newblock Noncommutative bilinear algorithms for 3 $\times$ 3 matrix
multiplication.
\newblock {\em SIAM J. Comp.}, 15(2):595--603, 1986.
\bibitem{Joslyn_2016}
C.~Joslyn, B.~Praggastis, E.~Purvine, A.~Sathanur, M.~Robinson, and
S.~Ranshous.
\newblock Local homology dimension as a network science measure.
\newblock In {\em SIAM Workshop on Network Science}, Boston, July 2016.
\bibitem{knuth1997art}
D.~E. Knuth.
\newblock {\em The Art of Computer Programming}, volume~3.
\newblock Pearson, 1997.
\bibitem{laderman1976noncommutative}
J.~D. Laderman.
\newblock A noncommutative algorithm for multiplying 3 $\times$ 3 matrices
using 23 multiplications.
\newblock {\em Bull. Amer. Math. Soc.}, 82(1):126--128, 1976.
\bibitem{landsberg2017geometry}
J.~M. Landsberg.
\newblock {\em Geometry and Complexity Theory}.
\newblock Cambridge, 2017.
\bibitem{Lee_2007}
J.-W. Lee, M.~Chiang, and R.~Calderbank.
\newblock Utility-optimal random-access control.
\newblock {\em IEEE Trans. Wireless Comm.}, 6(7):2741--2751, 2007.
\bibitem{makarov1986algorithm}
Oleg~M Makarov.
\newblock An algorithm for multiplying 3$\times$ 3 matrices.
\newblock {\em USSR Comp. Math. Math. Phys.}, 26(1):179--180, 1986.
\bibitem{McCabe1976}
T.~J. McCabe.
\newblock A complexity measure.
\newblock {\em IEEE Trans. Soft. Eng.}, SE-2:308, 1976.
\bibitem{Medeiros2017}
N.~Medeiros, N.~Ivaki, P.~Costa, and M.~Vieira.
\newblock Software metrics as indicators of security vulnerabilities.
\newblock In {\em ISSRE}, 2017.
\bibitem{Mesnard2016}
F.~Mesnard, \'E. Payet, and W.~Vanhoof.
\newblock Towards a framework for algorithm recognition in binary code.
\newblock In {\em PPDP}, 2016.
\bibitem{Nandagopal_2000}
T.~Nandagopal, T.-E. Kim, X.~Gao, and V.~Bharghavan.
\newblock Achieving {MAC} layer fairness in wireless packet networks.
\newblock In {\em MobiCom}, pages 87--–98, 2002.
\bibitem{nielson1992semantics}
H.~R. Nielson and F.~Nielson.
\newblock {\em Semantics with Applications}.
\newblock Springer, 1992.
\bibitem{Noubir_2004}
G.~Noubir.
\newblock On connectivity in \emph{ad hoc} networks under jamming using
directional antennas and mobility.
\newblock In {\em Wired/Wireless Internet Comm.}, pages 186--200. Springer,
2004.
\bibitem{Oudot2015}
S.~Y. Oudot.
\newblock {\em Persistence Theory: From Quiver Representations to Data
Analysis}.
\newblock AMS, 2015.
\bibitem{richardson1969some}
D.~Richardson.
\newblock Some undecidable problems involving elementary functions of a real
variable.
\newblock {\em J. Symb. Logic}, 33(4):514--520, 1969.
\bibitem{RobinsonGlobalSIP2014}
M.~Robinson.
\newblock Analyzing wireless communication network vulnerability with
homological invariants.
\newblock In {\em GlobalSIP}, 2014.
\bibitem{Robinson_2014}
M.~Robinson.
\newblock {\em Topological Signal Processing}.
\newblock Springer, 2014.
\bibitem{robinson2017sheaf}
M.~Robinson.
\newblock Sheaf and duality methods for analyzing multi-model systems.
\newblock In {\em Recent Applications of Harmonic Analysis to Function Spaces,
Differential Equations, and Data Science}, pages 653--703. Springer, 2017.
\bibitem{Scoville_2019}
N.~A. Scoville.
\newblock {\em Discrete Morse Theory}.
\newblock American Mathematical Society.
\bibitem{Shalaby2017}
M.~Shalaby, T.~Mehrez, A.~El-Mougy, K.~Abdulnasser, and A.~Al-Safty.
\newblock Automatic algorithm recognition of source-code using machine
learning.
\newblock In {\em ICMLA}, 2017.
\bibitem{strassen1969gaussian}
V.~Strassen.
\newblock Gaussian elimination is not optimal.
\newblock {\em Numerische Mathematik}, 13(4):354--356, 1969.
\bibitem{Taherkhani2010}
A.~Taherkhani, A.~Korhonen, and L.~Malmi.
\newblock Recognizing algorithms using language constructs, software metrics,
and roles of variables: an experiment with sorting algorithms.
\newblock {\em Computer J.}, 54:1049--1066, 2010.
\bibitem{wachs2007poset}
M.~L. Wachs.
\newblock Poset topology: tools and applications.
\newblock In E.~Miller, V.~Reiner, and B.~Sturmfels, editors, {\em Geometric
Combinatorics}.
\bibitem{winograd1971multiplication}
S.~Winograd.
\newblock On multiplication of 2 $\times$ 2 matrices.
\newblock {\em Linear Algebra Appl.}, 4(4):381--388, 1971.
\bibitem{Yang_2002}
X.~Yang and N.~Vaidya.
\newblock Priority scheduling in wireless \emph{ad hoc} networks.
\newblock In {\em MobiHoc}, 2002.
\bibitem{yanofsky2017galois}
N.~S. Yanofsky.
\newblock Galois theory of algorithms.
\newblock In {\em Rohit Parikh on Logic, Language and Society}, pages 323--347.
Springer, 2017.
\bibitem{zomorodian2005topology}
A.~J. Zomorodian.
\newblock {\em Topology for Computing}.
\newblock Cambridge, 2005.
\end{thebibliography}{10}10
\bibitem{nsnam}nsnam
The {NS-2}NS-2 network simulator.
\newblock {\tt http://www.nsnam.org/}\tt http://www.nsnam.org/.
\newblock Accessed: 2016-05-23.
\bibitem{Alves2016}Alves2016
H.~Alves, B.~Fonseca, and N.~Antunes.
\newblock Software metrics and security vulnerabilities: dataset and
exploratory study.
\newblock In {\em IDCC}\em IDCC, 2016.
\bibitem{ambrose2020topological}ambrose2020topological
K.~Ambrose, S.~Huntsman, M.~Robinson, and M.~Yutin.
\newblock Topological differential testing.
\newblock {\em arXiv preprint arXiv:2003.00976}\em arXiv preprint arXiv:2003.00976, 2020.
\bibitem{Arulselvan_2009}Arulselvan_2009
A.~Arulselvan, C.~W. Commander, L.~Elefteriadou, and P.~M. Pardalos.
\newblock Detecting critical nodes in sparse graphs.
\newblock {\em Comp. Operations Res.}\em Comp. Operations Res., 36(7):2193--2200, 2009.
\bibitem{Atkin1974}Atkin1974
R.~Atkin.
\newblock {\em Mathematical Structure in Human Affairs}\em Mathematical Structure in Human Affairs.
\newblock Heinemann, 1974.
\bibitem{ballard2019geometry}ballard2019geometry
G.~Ballard, C.~Ikenmeyer, J.M. Landsberg, and N.~Ryder.
\newblock The geometry of rank decompositions of matrix multiplication {II}II: 3
$\times$\times 3 matrices.
\newblock {\em J. Pure Appl. Algebra}\em J. Pure Appl. Algebra, 223(8):3205--3224, 2019.
\bibitem{barmak2011algebraic}barmak2011algebraic
J.~A. Barmak.
\newblock {\em Algebraic Topology of Finite Topological Spaces and
Applications}\em Algebraic Topology of Finite Topological Spaces and
Applications.
\newblock Springer, 2011.
\bibitem{berger2019equivalent}berger2019equivalent
G.~O. Berger, , P.-A. Absil, L.~{De Lathauwer}De Lathauwer, R.~M. Jungers, and M.~{Van
Barel}Van
Barel.
\newblock Equivalent polyadic decompositions of matrix multiplication tensors.
\newblock {\em arXiv preprint arXiv:1902.03950}\em arXiv preprint arXiv:1902.03950, 2019.
\bibitem{Carlsson2009}Carlsson2009
G.~Carlsson.
\newblock Topology and data.
\newblock {\em Bull. Amer. Math. Soc.}\em Bull. Amer. Math. Soc., 46:255--308, 2009.
\bibitem{Chiang_2007}Chiang_2007
M.~Chiang, S.~Low, R.~Calderbank, and J.~Doyle.
\newblock Layering as optimization decomposition: a mathematical theory of
network architectures.
\newblock {\em Proc. IEEE}\em Proc. IEEE, 95(1), January 2007.
\bibitem{chokaev2018two}chokaev2018two
B.~V. Chokaev and G.~N. Shumkin.
\newblock Two bilinear (3 $\times$\times 3)-matrix multiplication algorithms of
complexity 25.
\newblock {\em Moscow U. Comp. Math. Cyber.}\em Moscow U. Comp. Math. Cyber., 42(1):23--30, 2018.
\bibitem{Chowdhury2019}Chowdhury2019
S.~Chowdhury, T.~Gebhart, S.~Huntsman, and M.~Yutin.
\newblock Path homologies of deep feedforward networks.
\newblock In {\em ICMLA}\em ICMLA, 2019.
\bibitem{Chowdhury2018}Chowdhury2018
S.~Chowdhury and F.~M\'emoli.
\newblock A functorial {D}Dowker theorem and persistent homology of asymmetric
networks.
\newblock {\em J. Appl. Comp. Topology}\em J. Appl. Comp. Topology, 2(1):115, 2018.
\bibitem{courtois2011new}courtois2011new
N.~T. Courtois, G.~V. Bard, and D.~Hulme.
\newblock A new general-purpose method to multiply 3 $\times$\times 3 matrices using
only 23 multiplications.
\newblock {\em arXiv preprint arXiv:1108.2830}\em arXiv preprint arXiv:1108.2830, 2011.
\bibitem{DiSumma_2011}DiSumma_2011
M.~{Di Summa}Di Summa, A.~Grosso, and M.~Locatelli.
\newblock Complexity of the critical node problem over trees.
\newblock {\em Comp. Operations Res.}\em Comp. Operations Res., 38(12):1766--1774, 2011.
\bibitem{dinh2012}dinh2012
T.~N. Dinh, Y.~Xuan, M.~T. Thai, P.~M. Pardalos, and T.~Znati.
\newblock On new approaches of assessing network vulnerability: hardness and
approximation.
\newblock {\em IEEE/ACM Trans. Networking}\em IEEE/ACM Trans. Networking, 20(2):609--619, 2012.
\bibitem{Dowker1952}Dowker1952
C.~H. Dowker.
\newblock Homology groups of relations.
\newblock {\em Ann. Math.}\em Ann. Math., 56:84, 1952.
\bibitem{Du2019}Du2019
X.~Du, B.~Chen, Y.~Li, J.~Guo, Y.~Zhou, Y.~Liu, and Y.~Jiang.
\newblock {LEOPARD}LEOPARD: identifying vulnerable code for vulnerability assessment
through program metrics.
\newblock In {\em ICSE}\em ICSE, 2019.
\bibitem{Dullien2009}Dullien2009
T.~Dullien and S.~Porst.
\newblock {REIL}REIL: a platform-independent intermediate representation of
disassembled code for static code analysis.
\newblock In {\em CanSecWest}\em CanSecWest, 2009.
\bibitem{Duran2011}Duran2011
D.~Duran, D.~Weston, and M.~Miller.
\newblock Targeted taint driven fuzzing using software metrics.
\newblock In {\em CanSecWest}\em CanSecWest, 2011.
\bibitem{Eagle2011}Eagle2011
C.~Eagle.
\newblock {\em The IDA Pro Book: The Unofficial Guide to the World's Most
Popular Disassembler}\em The IDA Pro Book: The Unofficial Guide to the World's Most
Popular Disassembler.
\newblock No Starch Press, 2011.
\bibitem{Ebert2016}Ebert2016
C.~Ebert and J.~Cain.
\newblock Cyclomatic complexity.
\newblock {\em IEEE Soft.}\em IEEE Soft., 33:27, 2016.
\bibitem{Edelsbrunner2010}Edelsbrunner2010
H.~Edelsbrunner and J.~L. Harer.
\newblock {\em Computational Topology: An Introduction}\em Computational Topology: An Introduction.
\newblock AMS, 2010.
\bibitem{Erdmann2017}Erdmann2017
M.~Erdmann.
\newblock Topology of privacy: lattice structures and information bubbles for
inference and obfuscation.
\newblock {\em arXiv preprint arXiv:1712.04130}\em arXiv preprint arXiv:1712.04130, 2017.
\bibitem{fajstrup2016directed}fajstrup2016directed
L.~Fajstrup, E.~Goubault, E.~Haucourt, S.~Mimram, and M.~Raussen.
\newblock {\em Directed Algebraic Topology and Concurrency}\em Directed Algebraic Topology and Concurrency.
\newblock Springer, 2016.
\bibitem{Ghrist_2008}Ghrist_2008
R.~Ghrist.
\newblock Barcodes: the persistent topology of data.
\newblock {\em Bull. Amer. Math. Soc.}\em Bull. Amer. Math. Soc., 45(1):61, 2008.
\bibitem{GhristEAT}GhristEAT
R.~Ghrist.
\newblock {\em Elementary Applied Topology}\em Elementary Applied Topology.
\newblock Createspace, 2014.
\bibitem{Ghrist2012}Ghrist2012
R.~Ghrist, D.~Lipsky, J.~Derenick, and A.~Speranzon.
\newblock Topological landmark-based navigation and mapping.
\newblock {\em preprint}\em preprint, 2012.
\bibitem{Grigoryan2018}Grigoryan2018
A.~Grigor'yan, R.~Jimenez, Yu. Muranov, and S.-T. Yau.
\newblock On the path homology theory of digraphs and {Eilenberg-Steenrod}Eilenberg-Steenrod
axioms.
\newblock {\em Homology Homotopy Appl.}\em Homology Homotopy Appl., 20:179, 2018.
\bibitem{Grigoryan2018b}Grigoryan2018b
A.~Grigor'yan, Yu. Muranov, V.~Vershinin, and S.-T. Yau.
\newblock Path homology theory of multigraphs and quivers.
\newblock {\em Forum Math.}\em Forum Math., 30:1319, 2018.
\bibitem{Grigoryan2014}Grigoryan2014
A.~Grigor'yan, Yu. Muranov, and S.-T. Yau.
\newblock Graphs associated with simplicial complexes.
\newblock {\em Homology Homotopy Appl.}\em Homology Homotopy Appl., 16:295, 2014.
\bibitem{Grigoryan2017}Grigoryan2017
A.~Grigor'yan, Yu. Muranov, and S.-T. Yau.
\newblock Homologies of graphs and {K}K\"unneth formulas.
\newblock {\em Comm. Anal. Geom.}\em Comm. Anal. Geom., 25:969, 2017.
\bibitem{Grigoryan2012}Grigoryan2012
A.~Grigor'yan, L.~Yong, Yu. Muranov, and S.-T. Yau.
\newblock Homologies of path complexes and digraphs.
\newblock {\em arXiv preprint arXiv:1207.2834}\em arXiv preprint arXiv:1207.2834, 2012.
\bibitem{Grigoryan2014b}Grigoryan2014b
A.~Grigor'yan, L.~Yong, Yu. Muranov, and S.-T. Yau.
\newblock Homotopy theory for digraphs.
\newblock {\em Pure Appl. Math. Quart.}\em Pure Appl. Math. Quart., 10:619, 2014.
\bibitem{Grigoryan2015}Grigoryan2015
A.~Grigor'yan, L.~Yong, Yu. Muranov, and S.-T. Yau.
\newblock Cohomology of digraphs and (undirected) graphs.
\newblock {\em Asian J. Math.}\em Asian J. Math., 19:887, 2015.
\bibitem{Gueye_2010}Gueye_2010
A.~Gueye, J.~C. Walrand, and V.~Anantharam.
\newblock Design of network topology in an adversarial environment.
\newblock In {\em Decision and Game Theory for Security}\em Decision and Game Theory for Security, pages 1--20.
Springer, 2010.
\bibitem{Hatcher_2002}Hatcher_2002
A.~Hatcher.
\newblock {\em Algebraic Topology}\em Algebraic Topology.
\newblock Cambridge, 2002.
\bibitem{herlihy2014distributed}herlihy2014distributed
M.~Herlihy, D.~Kozlov, and S.~Rajsbaum.
\newblock {\em Distributed Computing Through Combinatorial Topology}\em Distributed Computing Through Combinatorial Topology.
\newblock Morgan Kaufmann, 2014.
\bibitem{heule2019new}heule2019new
M.~J.~H. Heule, M.~Kauers, and M.~Seidl.
\newblock New ways to multiply 3 $\times$\times 3-matrices.
\newblock {\em arXiv preprint arXiv:1905.10192}\em arXiv preprint arXiv:1905.10192, 2019.
\bibitem{Huntsman2018}Huntsman2018
S.~Huntsman.
\newblock Topological mixture estimation.
\newblock In {\em ICML}\em ICML, 2018.
\bibitem{Huntsman2020}Huntsman2020
S.~Huntsman.
\newblock Generalizing cyclomatic complexity via path homology.
\newblock {\em arXiv preprint arXiv:2003.00944}\em arXiv preprint arXiv:2003.00944, 2020.
\bibitem{Iozzo2010}Iozzo2010
V.~Iozzo.
\newblock 0-knowledge fuzzing.
\newblock In {\em Black Hat DC}\em Black Hat DC, 2010.
\bibitem{Jain_2003}Jain_2003
K.~Jain, J.~Padhye, V.~Padmanabhan, and L.~Qiu.
\newblock Impact of interference on multi-hop wireless network performance.
\newblock In {\em MobiCom}\em MobiCom, 2003.
\bibitem{johnson1986noncommutative}johnson1986noncommutative
R.~W. Johnson and A.~M. McLoughlin.
\newblock Noncommutative bilinear algorithms for 3 $\times$\times 3 matrix
multiplication.
\newblock {\em SIAM J. Comp.}\em SIAM J. Comp., 15(2):595--603, 1986.
\bibitem{Joslyn_2016}Joslyn_2016
C.~Joslyn, B.~Praggastis, E.~Purvine, A.~Sathanur, M.~Robinson, and
S.~Ranshous.
\newblock Local homology dimension as a network science measure.
\newblock In {\em SIAM Workshop on Network Science}\em SIAM Workshop on Network Science, Boston, July 2016.
\bibitem{knuth1997art}knuth1997art
D.~E. Knuth.
\newblock {\em The Art of Computer Programming}\em The Art of Computer Programming, volume~3.
\newblock Pearson, 1997.
\bibitem{laderman1976noncommutative}laderman1976noncommutative
J.~D. Laderman.
\newblock A noncommutative algorithm for multiplying 3 $\times$\times 3 matrices
using 23 multiplications.
\newblock {\em Bull. Amer. Math. Soc.}\em Bull. Amer. Math. Soc., 82(1):126--128, 1976.
\bibitem{landsberg2017geometry}landsberg2017geometry
J.~M. Landsberg.
\newblock {\em Geometry and Complexity Theory}\em Geometry and Complexity Theory.
\newblock Cambridge, 2017.
\bibitem{Lee_2007}Lee_2007
J.-W. Lee, M.~Chiang, and R.~Calderbank.
\newblock Utility-optimal random-access control.
\newblock {\em IEEE Trans. Wireless Comm.}\em IEEE Trans. Wireless Comm., 6(7):2741--2751, 2007.
\bibitem{makarov1986algorithm}makarov1986algorithm
Oleg~M Makarov.
\newblock An algorithm for multiplying 3$\times$\times 3 matrices.
\newblock {\em USSR Comp. Math. Math. Phys.}\em USSR Comp. Math. Math. Phys., 26(1):179--180, 1986.
\bibitem{McCabe1976}McCabe1976
T.~J. McCabe.
\newblock A complexity measure.
\newblock {\em IEEE Trans. Soft. Eng.}\em IEEE Trans. Soft. Eng., SE-2:308, 1976.
\bibitem{Medeiros2017}Medeiros2017
N.~Medeiros, N.~Ivaki, P.~Costa, and M.~Vieira.
\newblock Software metrics as indicators of security vulnerabilities.
\newblock In {\em ISSRE}\em ISSRE, 2017.
\bibitem{Mesnard2016}Mesnard2016
F.~Mesnard, \'E. Payet, and W.~Vanhoof.
\newblock Towards a framework for algorithm recognition in binary code.
\newblock In {\em PPDP}\em PPDP, 2016.
\bibitem{Nandagopal_2000}Nandagopal_2000
T.~Nandagopal, T.-E. Kim, X.~Gao, and V.~Bharghavan.
\newblock Achieving {MAC}MAC layer fairness in wireless packet networks.
\newblock In {\em MobiCom}\em MobiCom, pages 87--–98, 2002.
\bibitem{nielson1992semantics}nielson1992semantics
H.~R. Nielson and F.~Nielson.
\newblock {\em Semantics with Applications}\em Semantics with Applications.
\newblock Springer, 1992.
\bibitem{Noubir_2004}Noubir_2004
G.~Noubir.
\newblock On connectivity in \emph{ad hoc} networks under jamming using
directional antennas and mobility.
\newblock In {\em Wired/Wireless Internet Comm.}\em Wired/Wireless Internet Comm., pages 186--200. Springer,
2004.
\bibitem{Oudot2015}Oudot2015
S.~Y. Oudot.
\newblock {\em Persistence Theory: From Quiver Representations to Data
Analysis}\em Persistence Theory: From Quiver Representations to Data
Analysis.
\newblock AMS, 2015.
\bibitem{richardson1969some}richardson1969some
D.~Richardson.
\newblock Some undecidable problems involving elementary functions of a real
variable.
\newblock {\em J. Symb. Logic}\em J. Symb. Logic, 33(4):514--520, 1969.
\bibitem{RobinsonGlobalSIP2014}RobinsonGlobalSIP2014
M.~Robinson.
\newblock Analyzing wireless communication network vulnerability with
homological invariants.
\newblock In {\em GlobalSIP}\em GlobalSIP, 2014.
\bibitem{Robinson_2014}Robinson_2014
M.~Robinson.
\newblock {\em Topological Signal Processing}\em Topological Signal Processing.
\newblock Springer, 2014.
\bibitem{robinson2017sheaf}robinson2017sheaf
M.~Robinson.
\newblock Sheaf and duality methods for analyzing multi-model systems.
\newblock In {\em Recent Applications of Harmonic Analysis to Function Spaces,
Differential Equations, and Data Science}\em Recent Applications of Harmonic Analysis to Function Spaces,
Differential Equations, and Data Science, pages 653--703. Springer, 2017.
\bibitem{Scoville_2019}Scoville_2019
N.~A. Scoville.
\newblock {\em Discrete Morse Theory}\em Discrete Morse Theory.
\newblock American Mathematical Society.
\bibitem{Shalaby2017}Shalaby2017
M.~Shalaby, T.~Mehrez, A.~El-Mougy, K.~Abdulnasser, and A.~Al-Safty.
\newblock Automatic algorithm recognition of source-code using machine
learning.
\newblock In {\em ICMLA}\em ICMLA, 2017.
\bibitem{strassen1969gaussian}strassen1969gaussian
V.~Strassen.
\newblock Gaussian elimination is not optimal.
\newblock {\em Numerische Mathematik}\em Numerische Mathematik, 13(4):354--356, 1969.
\bibitem{Taherkhani2010}Taherkhani2010
A.~Taherkhani, A.~Korhonen, and L.~Malmi.
\newblock Recognizing algorithms using language constructs, software metrics,
and roles of variables: an experiment with sorting algorithms.
\newblock {\em Computer J.}\em Computer J., 54:1049--1066, 2010.
\bibitem{wachs2007poset}wachs2007poset
M.~L. Wachs.
\newblock Poset topology: tools and applications.
\newblock In E.~Miller, V.~Reiner, and B.~Sturmfels, editors, {\em Geometric
Combinatorics}\em Geometric
Combinatorics.
\bibitem{winograd1971multiplication}winograd1971multiplication
S.~Winograd.
\newblock On multiplication of 2 $\times$\times 2 matrices.
\newblock {\em Linear Algebra Appl.}\em Linear Algebra Appl., 4(4):381--388, 1971.
\bibitem{Yang_2002}Yang_2002
X.~Yang and N.~Vaidya.
\newblock Priority scheduling in wireless \emph{ad hoc} networks.
\newblock In {\em MobiHoc}\em MobiHoc, 2002.
\bibitem{yanofsky2017galois}yanofsky2017galois
N.~S. Yanofsky.
\newblock Galois theory of algorithms.
\newblock In {\em Rohit Parikh on Logic, Language and Society}\em Rohit Parikh on Logic, Language and Society, pages 323--347.
Springer, 2017.
\bibitem{zomorodian2005topology}zomorodian2005topology
A.~J. Zomorodian.
\newblock {\em Topology for Computing}\em Topology for Computing.
\newblock Cambridge, 2005.
|
Conclusion
| false
|
2008.03299
| 5
|
End of preview. Expand
in Data Studio
README.md exists but content is empty.
- Downloads last month
- 19