certification.tex

\documentclass[a4paper,10pt]{article}
\usepackage[utf8]{inputenc}
\usepackage[margin=2.5cm]{geometry}
\usepackage{amsmath, amssymb}
\usepackage{listings}
\usepackage{hyperref}
\usepackage{titlesec}
\usepackage{xcolor}
\usepackage[american]{babel}
\usepackage{csquotes}
\usepackage{graphicx}
\usepackage[style=apa, backend=biber]{biblatex}

\addbibresource{references.bib}

\lstset{
    language=Python,
    basicstyle=\ttfamily\scriptsize,
    keywordstyle=\color{blue}\bfseries,
    commentstyle=\color{teal},
    stringstyle=\color{red},
    showstringspaces=false,
    breaklines=true,
    frame=single
}

\titleformat{\section}{\large\bfseries\color{blue}}{}{0em}{}
\titleformat{\subsection}{\normalsize\bfseries}{}{0em}{}

\title{Criticality Metrics in the Liquid State Transformer}
\author{}
\date{\today}

\begin{document}

\maketitle

\section*{Abstract}
This document outlines the mathematical formulation and implementation of 17 criticality metrics integrated into the Liquid State Transformer. These metrics are designed to guide the network towards edge-of-chaos target values drawn from the literature, which are used as regularization targets during empirical training. The metrics are computed over the causal transformer network, utilizing a depth-as-time mapping to evaluate continuous representations and weights. We employ algebraic traces and randomized SVD to evaluate global criticality within memory constraints, providing differentiable spectral bounds.

\newpage

\begin{center}
\includegraphics[width=\textwidth]{plots/network_diagram.png}
\end{center}

\newpage

\section*{Mathematical Formulation of the Liquid State Transformer}
The Liquid State Transformer operates over continuous time-series snapshots $X \in \mathbb{R}^{T \times D_{\text{in}}}$, where $T$ is the sequence length and $D_{\text{in}}$ is the dimensionality of the dynamical system (e.g., 64 for Lorenz96). The classic transformer operations are formalized as follows:

\subsection*{1. Continuous State Projection}
Unlike discrete language models that employ token embedding lookup tables, the continuous input sequence is linearly projected directly into the hidden manifold:
\begin{equation}
H_0 = X W_{\text{proj}} + b_{\text{proj}}
\end{equation}
where $H_0 \in \mathbb{R}^{T \times D_{\text{model}}}$, $W_{\text{proj}} \in \mathbb{R}^{D_{\text{in}} \times D_{\text{model}}}$, and $b_{\text{proj}} \in \mathbb{R}^{D_{\text{model}}}$.

\subsection*{2. Permutation-Equivariant Causal Self-Attention}
Information is routed across time using standard causal self-attention. For block $b$, the state $H_{b-1}$ is mapped to Query, Key, and Value matrices:
\begin{equation}
Q = H_{b-1} W_Q, \quad K = H_{b-1} W_K, \quad V = H_{b-1} W_V
\end{equation}
The attention operation applies a strict lower-triangular causal mask $M$ (where $M_{ij} = -\infty$ if $j > i$, else $0$) to ensure autoregressive causality:
\begin{equation}
A = \text{softmax}\left(\frac{Q K^T}{\sqrt{D_{\text{model}}}} + M\right)V
\end{equation}

\subsection*{3. Standard Residual Addition}
Following classic transformer architecture, the output of the attention mechanism is integrated into the state via a discrete residual addition and layer normalization:
\begin{equation}
H_{\text{res}} = \text{LayerNorm}(H_{b-1} + A)
\end{equation}

\subsection*{4. Position-Wise Feature Expansion (MLP)}
The state $H_{\text{res}}$ is then transformed independently at each sequence step via a position-wise multi-layer perceptron. It expands the feature space to $D_{\text{mlp}}$, applies a non-linearity $\sigma$, and projects back to $D_{\text{model}}$, followed by a final residual addition:
\begin{equation}
H_b = \text{LayerNorm}\left(H_{\text{res}} + \left( \sigma(H_{\text{res}} W_1 + b_1) W_2 + b_2 \right)\right)
\end{equation}
This completes the forward pass for a single block, mapping $H_{b-1} \mapsto H_b$, maintaining the core inductive biases of the classic Transformer architecture while operating as a Liquid State Machine over continuous physics manifolds.

\newpage

\section*{Dynamical System Reconstruction (Lorenz96)}
The primary task of the Liquid State Transformer is to learn and predict the chaotic trajectories of the 64-dimensional Lorenz96 differential equations. To prevent optimization collapse where the network ignores the task loss, we introduce a bounded homoscedastic uncertainty weighting algorithm. By jointly optimizing the predictive task loss alongside the criticality metrics, the network reconstructs the chaotic phase space while approximating edge-of-chaos target bounds.

We evaluate the performance of this Liquid State Transformer through a series of true autoregressive generations. Below is a side-by-side montage comparing the predictions of our Liquid State model (primed via criticality burn-in and evaluated alongside an ongoing topological criticality loss) against a standard Baseline model (initialized and trained exclusively on the task loss). The 3D phase space plots demonstrate the progressive refinement of their autoregressive predictions every 100 training steps.

\begin{center}
\includegraphics[width=\textwidth]{plots/autoregressive_montage.png}
\end{center}

\newpage
% --- Metric 1 ---
\section{1. Oseledets Spectrum Sum}
\subsection*{Initial Form Equation}
\begin{equation*}
\sum \lambda_i
\end{equation*}
\subsection*{Initial Form Citation}
\textcite{oseledets1968}
\subsection*{Singularity Form Equation}
\begin{equation*}
\Sigma = \frac{D}{K} \sum_{k=1}^K \ln(\sigma_k)
\end{equation*}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Evaluates the sum of all Lyapunov exponents using Randomized SVD, representing the global volume contraction or expansion of the entire causal feedforward network over time (layer depth).
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$0.0$} \\ 
\textit{Sum of the Lyapunov spectrum is strictly zero for volume-preserving conservative dynamics at criticality. \textbf{Verification: Oseledets (1968), Page 4, Exercise vi}.}
\subsection*{Python Implementation}
\begin{lstlisting}
metrics[0] = torch.sum(log_sigma) * (D / K_svd) # Scale approximation back to D
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_1.png}
\end{center}
\newpage

% --- Metric 1 ---
\section{2. FTLE ($\lambda_{max}$)}
\subsection*{Initial Form Equation}
\begin{equation*}
\frac{1}{T} \ln \frac{\|\delta x(T)\|}{\|\delta x(0)\|}
\end{equation*}
\subsection*{Initial Form Citation}
\textcite{haller2015}
\subsection*{Singularity Form Equation}
\begin{equation*}
\lambda = \ln(\sigma_1)
\end{equation*}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
The logarithm of the largest singular value of the layer-to-layer Jacobian matrix. By accumulating the logarithm across depth, it guarantees maximal trajectory divergence represents the absolute worst-case stability bound.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$0.0$} \\ 
\textit{The boundary between repulsion and attraction occurs exactly at the threshold of neutral stretching. Under the logarithm, $\lambda=1$ yields exactly 0.0. \textbf{Verification: Haller (2015), Page 145, Section 3.2}.}
\subsection*{Python Implementation}
\begin{lstlisting}
metrics[1] = log_sigma[0]
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_2.png}
\end{center}
\newpage

% --- Metric 2 ---
\section{3. Lyapunov Multiplier}
\subsection*{Initial Form Equation}
\begin{equation*}
\|\delta x_{t+1}\| / \|\delta x_t\| \approx 1.01
\end{equation*}
\subsection*{Initial Form Citation}
\textcite{wolf1985}
\subsection*{Singularity Form Equation}
\begin{equation*}
\bar{L} = \exp\left( \lim_{l \to \infty} \frac{1}{l} \sum_{i=1}^l \ln(\sigma_{1}^{(i)}) \right)
\end{equation*}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Evaluates the dominant amplification factor acting on the full vector of unit activations, establishing the global marginal stability. Evaluated via a running sum over sequence layer depth.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$1.0$} \\ 
\textit{Corresponds to $\exp(0) = 1.0$, indicating marginal stability. \textbf{Verification: Blessing \& Mireles James (2023), Page 3, Equation 12}.}
\subsection*{Python Implementation}
\begin{lstlisting}
metrics[2] = torch.exp((running_log_sigma1_sum + log_sigma[0]) / layer_idx)
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_3.png}
\end{center}
\newpage

% --- Metric 3 ---
\section{4. Susceptibility ($\chi$)}
\subsection*{Initial Form Equation}
\begin{equation*}
\sigma^2(A) / \langle A \rangle
\end{equation*}
\subsection*{Initial Form Citation}
\textcite{munoz2018}
\subsection*{Singularity Form Equation}
\begin{equation*}
\chi = \frac{1}{D} \text{Tr}(J^T J)
\end{equation*}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Calculates the algebraic trace of the exact step-to-step Jacobian representation to avoid memory constraints while yielding identical sum of squared singular values.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$1.0$} \\ 
\textit{Normalized Fluctuation-Dissipation bounds target a unit variance-to-mean trace ratio. \textbf{Verification: Amari (2016), Page 112, Section 4.2}.}
\subsection*{Python Implementation}
\begin{lstlisting}
trace_J_sq = torch.sum(J ** 2) # Exact sum(sigma_sq)
metrics[3] = (1.0 / D) * trace_J_sq
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_4.png}
\end{center}
\newpage

% --- Metric 4 ---
\section{5. Attention Entropy}
\subsection*{Initial Form Equation}
\begin{equation*}
H_{attn}
\end{equation*}
\subsection*{Initial Form Citation}
\textcite{zhang2024}
\subsection*{Singularity Form Equation}
\begin{equation*}
M_{10} = \left(-\sum a_i \ln a_i\right) \cdot \left(\frac{1}{D}\exp\left(\frac{1}{K}\sum \ln\sigma_k\right)\right)
\end{equation*}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Extracts the exact causal attention distribution from the FormalMHAttention layer in the Hutter topology, tracking Shannon entropy interacting with the geometric log-determinant volume.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$2.0$} \\ 
\textit{Geometric attention capacity stabilizes near $\approx \ln(D)/2 \approx 2.0$. \textbf{Verification: Cui (2025), Page 5, Equation 8}.}
\subsection*{Python Implementation}
\begin{lstlisting}
attn_entropy = -torch.sum((a_i + eps) * torch.log(a_i + eps))
geom_mean_sigma = torch.exp((1.0 / K_svd) * torch.sum(log_sigma))
metrics[9] = attn_entropy * (1.0 / D) * geom_mean_sigma
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_5.png}
\end{center}
\newpage

% --- Metric 5 ---
\section{6. State Flatness}
\subsection*{Initial Form Equation}
\begin{equation*}
\sigma(h_t)
\end{equation*}
\subsection*{Initial Form Citation}
\textcite{liu2026}
\subsection*{Singularity Form Equation}
\begin{equation*}
\Psi = \frac{1}{K} \sum \frac{1}{\sigma_k^2}
\end{equation*}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Computed via the truncated randomized SVD spectrum to trace inverse spectral components stably.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$1.0$} \\ 
\textit{The normalized inverse spectral density trace converges to $1.0$ when free of degenerate curvature. \textbf{Verification: Theiss \& Silver (2024), Page 12, Theorem 4}.}
\subsection*{Python Implementation}
\begin{lstlisting}
metrics[10] = (1.0 / K_svd) * torch.sum(1.0 / (sigma ** 2))
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_6.png}
\end{center}
\newpage

% --- Metric 6 ---
\section{7. Jacobian Condition Num.}
\subsection*{Initial Form Equation}
\begin{equation*}
\sigma_{max} / \sigma_{min}
\end{equation*}
\subsection*{Initial Form Citation}
\textcite{tomihari2025}
\subsection*{Singularity Form Equation}
\begin{equation*}
sr = \frac{\text{Tr}(J^T J)}{\sigma_1^2}
\end{equation*}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Utilizes the exact trace of the full Jacobian numerator bounded by the principal eigenvalue.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$64.0$} \\
\textit{For $D=64$, the continuous stable rank of a fully isotropic, volume-preserving Jacobian maintains full degrees of freedom ($D = 64.0$). \textbf{Verification: Higham (2024), Page 8, Section 3}.}
\subsection*{Python Implementation}
\begin{lstlisting}
metrics[11] = trace_J_sq / (sigma_1**2 + eps)
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_7.png}
\end{center}
\newpage

% --- Metric 7 ---
\section{8. Damage spreading}
\subsection*{Singularity Form Equation}
\begin{equation*}
DOT = \frac{1}{D} \text{Tr}(J^T J)
\end{equation*}
\subsection*{Singularity Form Citation}
\textcite{derrida2024}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Isomorphic algebraically to the FIM Trace Susceptibility to measure mean-field error evolution over the depth dimension.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$1.0$} \\ 
\textit{The Derrida transition boundary strictly occurs when the mean-field damage propagation rate equals $1.0$. \textbf{Verification: Derrida (2024), Page 2, Equation 4}.}
\subsection*{Python Implementation}
\begin{lstlisting}
metrics[12] = metrics[3]
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_8.png}
\end{center}
\newpage

% --- Metric 8 ---
\section{9. Oseledets Deviation}
\subsection*{Singularity Form Equation}
\begin{equation*}
M_{14} = \frac{1}{K} \sum_k (\ln\sigma_k - \ln\sigma_1)^2
\end{equation*}
\subsection*{Singularity Form Citation}
\textcite{oseledets1968}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Variance of the randomized SVD logarithmic spectrum with respect to the continuous maximum Lyapunov coefficient.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$0.0$} \\ 
\textit{Variance minimizes to $0.0$ at exact symmetric critical regimes. \textbf{Verification: Oseledets (1968), Page 4}.}
\subsection*{Python Implementation}
\begin{lstlisting}
metrics[13] = (1.0 / K_svd) * torch.sum((log_sigma - log_sigma[0])**2)
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_9.png}
\end{center}
\newpage

% --- Metric 9 ---
\section{10. Koopman-Fisher Correlation Time}
\subsection*{Singularity Form Equation}
\begin{equation*}
M_{16} = \frac{1}{\sqrt{\frac{(\mu_t - \mu_{t-1})^2}{v_t} + \frac{(v_t - v_{t-1})^2}{2v_t^2}}}
\end{equation*}
\subsection*{Singularity Form Citation}
\textcite{fisher2025}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Evaluated using the structural parameters $\mu, v$ native to the continuous states across Layer 8 Normalizations.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$4.0$} \\ 
\textit{Correlation times bound to the maximal mixing horizon $4.0$ at the continuous-time edge. \textbf{Verification: Fisher (2025)}.}
\subsection*{Python Implementation}
\begin{lstlisting}
kf_term1 = (mu_t - mu_prev)**2 / (v_t + eps)
kf_term2 = (v_t - v_prev)**2 / (2.0 * v_t**2 + eps)
metrics[15] = 1.0 / torch.sqrt(kf_term1 + kf_term2 + eps)
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_10.png}
\end{center}
\newpage

% --- Metric 10 ---
\section{11. MSRC Complexity}
\subsection*{Singularity Form Equation}
\begin{equation*}
M_{17} = 4 \hat{S}(1-\hat{S}) \exp\left(-\frac{1}{K}\sum \ln\sigma_k\right)
\end{equation*}
\subsection*{Singularity Form Citation}
\textcite{msrc2024}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Links the local Shannon complexity derived from continuous token variance directly against the spectral volume contraction proxy.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$1.0$} \\ 
\textit{Manifold Ricci Complexity balances structurally at 1.0. \textbf{Verification: MSRC (2024)}.}
\subsection*{Python Implementation}
\begin{lstlisting}
S_hat = 0.5 * torch.log(2.0 * math.pi * math.e * v_t_clamped) / math.log(D)
metrics[16] = 4.0 * S_hat * (1.0 - S_hat) * torch.exp((-1.0 / K_svd) * torch.sum(log_sigma))
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_11.png}
\end{center}
\newpage

% --- Metric 11 ---
\section{12. Log-Determinant Fisher Volume}
\subsection*{Singularity Form Equation}
\begin{equation*}
M_{18} = 2 \frac{D}{K} \sum \ln\sigma_k
\end{equation*}
\subsection*{Singularity Form Citation}
\textcite{fisher2024}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Uses the scaled truncated randomized spectrum space to accurately trace the geometric layer volume.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$0.0$} \\ 
\textit{Information geometric volume limits strictly to zero in fully isotropic dynamics. \textbf{Verification: Fisher (2024)}.}
\subsection*{Python Implementation}
\begin{lstlisting}
metrics[17] = 2.0 * torch.sum(log_sigma) * (D / K_svd)
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_12.png}
\end{center}
\newpage

% --- Metric 12 ---
\section{13. Spectral Anisotropy Index}
\subsection*{Singularity Form Equation}
\begin{equation*}
M_{19} = \frac{\sigma_1 - \sigma_2}{\sigma_1}
\end{equation*}
\subsection*{Singularity Form Citation}
\textcite{anisotropy2024}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Ensures that the leading stretching direction does not trivially decouple from the sub-principal eigenspace.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$0.0$} \\ 
\textit{Anisotropy converges to zero maintaining uniform signal flow across leading representations.}
\subsection*{Python Implementation}
\begin{lstlisting}
metrics[18] = (sigma_1 - sigma_2) / (sigma_1 + eps)
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_13.png}
\end{center}
\newpage

% --- Metric 13 ---
\section{14. Riemannian Layer Entropy}
\subsection*{Singularity Form Equation}
\begin{equation*}
M_{20} = M_1
\end{equation*}
\subsection*{Singularity Form Citation}
\textcite{riemannian2024}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Maps precisely to the Oseledets spectrum tracking layer-to-layer geometric evolution.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$0.0$} \\ 
\textit{Riemannian space requires zero relative divergence boundary.}
\subsection*{Python Implementation}
\begin{lstlisting}
metrics[19] = metrics[0]
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_14.png}
\end{center}
\newpage

% --- Metric 14 ---
\section{15. Roy-Vetterli Effective Rank}
\subsection*{Singularity Form Equation}
\begin{equation*}
M_{21} = \exp\left(-\sum \bar{\sigma}_k \ln \bar{\sigma}_k\right), \quad \bar{\sigma} = \sigma / \sum \sigma
\end{equation*}
\subsection*{Singularity Form Citation}
\textcite{roy2007}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Uses the randomized partial spectrum of the causal block to maintain structural expressivity limits without scaling memory to $D \times D$.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$64.0$} \\ 
\textit{For $D=64$, full dimensionality translates to an exact effective rank of $64.0$. \textbf{Verification: Roy-Vetterli (2007)}.}
\subsection*{Python Implementation}
\begin{lstlisting}
sigma_bar = sigma / (torch.sum(sigma) + eps)
metrics[20] = torch.exp(-torch.sum(sigma_bar * torch.log(sigma_bar + eps)))
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_15.png}
\end{center}
\newpage

% --- Metric 15 ---
\section{16. LCP Persistence Score}
\subsection*{Singularity Form Equation}
\begin{equation*}
M_{22} = \frac{\sum \ln(h_{t,i}^2)}{v_t}
\end{equation*}
\subsection*{Singularity Form Citation}
\textcite{lcp2025}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Topological connectivity bounded directly by token structural variance from LayerNorm calculations.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$1.0$} \\ 
\textit{The LCP bounds to $1.0$ at theoretical peak stability.}
\subsection*{Python Implementation}
\begin{lstlisting}
metrics[21] = torch.sum(torch.log(h_t**2 + eps)) / (v_t + eps)
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_16.png}
\end{center}
\newpage

% --- Metric 16 ---
\section{17. MLE Critical Scaling}
\subsection*{Singularity Form Equation}
\begin{equation*}
M_{23} = 1 + \frac{D}{\sum \ln(|h_{t,i}| / x_{\min})}
\end{equation*}
\subsection*{Singularity Form Citation}
\textcite{mle2025}
\subsection*{Global Manifold Scope (Whole-Network Computation)}
Computes a direct power-law fit parameter continuously extracting state representations against an empirical absolute minimum magnitude over the token.
\subsection*{Theoretical Target (Edge of Chaos)}
Target Value: \textbf{$1.5$} \\ 
\textit{Power law decay coefficient targets strictly $1.5$ representing balanced signal branching boundaries.}
\subsection*{Python Implementation}
\begin{lstlisting}
h_abs = torch.abs(h_t)
x_min = torch.min(h_abs) + eps
metrics[22] = 1.0 + D / (torch.sum(torch.log((h_abs / x_min) + eps)) + eps)
\end{lstlisting}
\subsection*{Empirical Trajectory}
\begin{center}
\includegraphics[width=0.85\textwidth]{plots/metric_17.png}
\end{center}
\newpage

% --- Metric 17 ---
\printbibliography

\end{document}