Computational Physics Blog RSS Feed

Voxel engine for Markov algorithms

Sun, 11 Jan 2026 19:22:00 GMT

This is the result of project made for a Polytechnique course on 3D programming. It is a collobaration between me and Victor Llorca.

The code is available here: https://github.com/Jaymun723/Voxel-engine-for-Markov-algorithms.

To move the car use the WASD commands.

Introduction

In this project, we chose to implement procedural generation using a voxel (volumetric pixel) management system and generation rules derived from Markov algorithms.

To showcase this project, we placed this scene in an urban context, where a car can move along roads bordering plots generated using Markov algorithms.

1 - The Voxel Engine

A voxel engine is a method for displaying landscapes made up of blocks. The three main elements of a voxel engine are:

1.a) The blocks (or voxels)

In our project, the world is entirely filled with blocks. A block is simply a memory representation containing a type that defines how the space it occupies should be filled. The block type is also used by the Markov algorithm. The most important type is BlockType_Empty; it designates the absence of "physical" filling.

1.b) The chunks

A chunk is a grouping of (l * w * h) blocks. In our final version, a chunk is a group of 16 _ 16 _ 16 = 4,096 blocks. It is within the chunk that the first graphical phase occurs. Each chunk is assigned a mesh which will then be loaded or not. The mesh calculation involves a method called "Face culling" to avoid including unnecessary triangles. The principle relies on the fact that if two (non-empty) blocks share the same face, then that face is not visible to the user.

1.c) The chunk manager

The chunk manager is responsible for creating chunks as the player moves and keeping them in memory. If chunks are too far away (relative to a certain RENDER_DISTANCE), it unloads them to prevent the graphics card from becoming overloaded. To avoid conflicts during random generation, chunks currently being generated by the Markov method are not rendered but continue to be updated if they go outside the RENDER_DISTANCE.

2 - Markov Algorithms

The simplest structure used by Markov algorithms is the notion of a rule. A rule is simply defined by an application condition and an application function. It allows for the transformation of a set of voxels into another. For example, to start the pillars of a house, we wish to replace the pattern on the left with the pattern on the right. Concretely, we place a white pillar under a gray block that has only 2 neighbors.

A set of rules is called a Markov rule. It is applied in the following way:

Quantum Monte Carlo using World-line approach.

Sat, 20 Dec 2025 19:00:00 GMT

As part of my curriculum at Polytechnique I took a numerical physics class where we had to implement a research paper on a specific algorithms from computational physics.

I paired with a friend to implement a paper written by Assaad and Evertz on World-line and Determinantal Quantum Monte Carlo Methods for Spins, Phonos and Electrons.

The code is available here https://github.com/jaymun723/quantum-monte-carlo. You can find the PDF version of this article here.

Introduction

Problems in solid-state physics often require computationally demanding calculations, as their complexity typically scales as a power of system size. Spin chains and electron chains exhibit strong correlations, making exact solutions infeasible for large systems. However, one-dimensional (1D) spin chains provide a good balance between analytical tractability and physical relevance. In this study, the focus is placed on 1D XXZ chains, where correlations along the $x$ and $y$ axes are identical. This framework allows verification of theoretical predictions, assessment of approximations, and extrapolation of system properties to larger sizes. The Monte Carlo method described in Assaad and Evertz was implemented using three distinct update schemes.

We want to understand the physical properties of a system described by the following hamiltonian :

The XXZ Hamiltonian

The XXZ model describes a chain of interacting spin- $\frac{1}{2}$ particles with nearest-neighbour coupling and anisotropy between the $xy$ -plane and the $z$ -direction. The Hamiltonian is given by

\begin{equation} H = J_{x}\sum_{i}(S_{i}^{x}S_{i+1}^{x} + S_{i}^{y}S_{i+1}^{y}) + J_{z}\sum_{i}S_{i}^{z}S_{i+1}^{z} \end{equation}

Periodic boundary conditions are imposed such that $S_i = S_{i+n}$ . The coupling constants $J_x$ and $J_z$ define the physical regime of the system: the $J_z$ term represents an Ising-like interaction, favouring spin alignment ( $J_z<0$ ) or anti-alignment ( $J_z>0$ ), whereas $J_x$ induces quantum fluctuations through spin exchanges between neighbouring sites.

Computational complexity

Although the Hamiltonian can be written compactly, evaluating thermodynamic quantities such as specific heat, magnetic susceptibility, or spin stiffness requires computation of the partition function $Z = \mathrm{Tr}[e^{-\beta H}]$ .

This presents a mathematical challenge: the Hilbert space of a quantum system grows exponentially with its physical size. Since each local spin has two basis states ( $\ket{\uparrow}$ and $\ket{\downarrow}$ ), the dimension of the total Hilbert space is:

\begin{equation} \text{dim}(\mathcal{H}_n) = 2^n \end{equation}

For very small systems, one can perform diagonalization of the Hamiltonian matrix. However, this approach becomes impossible for large systems. Therefore, it is necessary to use other numerical approaches to solve for larger systems that do not require the full matrix diagonalization. Nevertheless, this exact calculation is crucial for testing the numerical method on smaller systems and for extrapolating to larger ones.

World-line approach

Trotterization

Following the derivation by Assaad and Evertz, the problem is simplified to make it tractable numerically. Let $H^{(i)} = J_{x}(S_{i}^{x}S_{i+1}^{x} + S_{i}^{y}S_{i+1}^{y}) + J_{z}S_{i}^{z}S_{i+1}^{z}$ be the two-site Hamiltonian, whose eigenvalues are known. Then, $H_1$ , the sum of the even two-sites, and $H_2$ , the sum of the odd terms, are two sums of commuting Hamiltonians. Introducing the discretisation $\Delta\tau \equiv \frac{\beta}{m}$ , Trotter’s formula approximates the partition function as:

\begin{align*} Z &= \text{Tr}[(e^{-\Delta\tau H})^{m}] \\ &= \text{Tr}[(e^{-\Delta\tau H_{1}}e^{-\Delta\tau H_{2}})^{m}] + \mathcal{O}(\Delta\tau^{2}) \\ &=\sum_{\pmb{\sigma}_1 \dots \pmb{\sigma}_{2m}} \bra{\pmb{\sigma}_1} e^{-\Delta\tau H_1} \ket{\pmb{\sigma}_{2m}} \cdots \braket{\pmb{\sigma}_2|e^{-\Delta\tau H_2}|\pmb{\sigma}_1} + \mathcal{O}(\Delta\tau^{2}) \end{align*}

Then, $\Delta\tau$ can then be interpreted as a discrete imaginary time step, mapping the 1D system into a 2D configuration with a progression from $\tau = 0$ to $\tau = m\Delta\tau = \beta$ as the additional dimension. This conversion allows to treat this system with a classical approach, thus enabling Monte Carlo methods.

Figure 1: World-line configuration for $8$ spin sites and $m=4$ . Red represents up spins and blue represents down spins.

This system can then be represented as a world-line configuration tracking the evolution of the chain's spin configurations through time and space. This representation possesses interesting properties, such as spin conservation, which simplifies the calculation by eliminating spin evolution or 2D configurations not having a meaningful contribution at first order on $\Delta\tau$ .

Periodic boundary conditions are chosen as they ensure that boundary effects are minimised. In practice, $m$ needs to be large enough so as to minimise the effects of the Trotterization.

As for the code, a Worldline is represented by a matrix of size $2m \times n$ containing $\pm1$ values depending on the state of the spin at that location.

Weight and energy of world-lines

Each Worldline has its own weight denoted as $\Omega(\omega)$ .

From Assaad and Evertz:

\begin{equation} \Omega(\omega) = \prod_{i=1}^{2m} \braket{\pmb{\sigma}_{i+1}|e^{-\Delta\tau H_i}|\pmb{\sigma}_{i}} \end{equation}

Where $H_i$ is $H_1$ if $i$ is odd and $H_2$ otherwise. Periodic boundary conditions give $\pmb{\sigma}_{2m+1}=\pmb{\sigma}_{1}$ .

With equation 3, the total weight is the product of the weight of each line.

\begin{equation} \braket{\pmb{\sigma}_{i+1}|e^{-\Delta\tau H_i}|\pmb{\sigma}_i} = \prod_{j=i}^{n/2} \braket{\sigma_{2j, i+1}, \sigma_{2j+1, i+1} |e^{-\Delta\tau H_i}|\sigma_{2j,i},\sigma_{2j+1,i}} \end{equation}

And equation 4 gives that each line weight is the product of the weight of each of its white plaquette.

Note: In the code i and j are switched from the usual convention to match the notation used in the paper.

The combination of equations 3 and 4 gives a straightforward algorithm to compute the weight of a Worldline:

\begin{equation} \Omega(\omega) = \prod_{p} W_p \end{equation}

Where $p$ is taken over all white plaquettes of $\omega$ and $W_p$ is the plaquette weight.

The energy associated with a configuration is: (from Assaad and Evertz)

\begin{equation} E(w) = -\frac{1}{m} \sum_{p} \frac{\partial}{\partial \Delta\tau} \ln(W_p) \end{equation}

Based on the weights provided in Assaad and Evertz:

Full Squares: $\begin{align} W &= e^{-\Delta\tau J_z / 4} \\ -\frac{\partial}{\partial \Delta\tau} \ln(W) &= -\frac{\partial}{\partial \Delta\tau} \left( -\frac{\Delta\tau J_z}{4} \right) = \frac{J_z}{4} \end{align}$
Vertical World-Lines: $\begin{align} W &= e^{\Delta\tau J_z / 4} \cosh\left(\frac{\Delta\tau J_x}{2}\right) \\ -\frac{\partial}{\partial \Delta\tau} \ln(W) &= -\left( \frac{J_z}{4} + \frac{J_x}{2} \tanh\left(\frac{\Delta\tau J_x}{2}\right) \right) \end{align}$
Crossed World-Lines: $\begin{align} W &= -e^{\Delta\tau J_z / 4} \sinh\left(\frac{\Delta\tau J_x}{2}\right) \\ -\frac{\partial}{\partial \Delta\tau} \ln|W| &= -\left( \frac{J_z}{4} + \frac{J_x}{2} \coth\left(\frac{\Delta\tau J_x}{2}\right) \right) \end{align}$

Generating world-line configurations

In order to test the validity of Trotter's formula and our calculation method, one can generate all possible valid configurations for small grids. This allows to make exact calculations and compare them to our results.

The first approach took was to generate all the spin configurations possible and to keep only the valid ones. While straightforward to implement, this approach was computationally inefficient, as it required computing $2^{2 m n}$ configurations.

The second approach consisted of recursively generating a world-line with backtracking whenever an invalid configuration was encountered. Enforcing periodic boundary conditions proved more difficult than expected, resulting in reduced implementation efficiency. Each world-line was generated until an invalid configuration appeared on a white plaquette, at which point the process was halted. However, this algorithm did not yield significant improvement over the previous one, as the number of configurations continued to grow exponentially. The method could be optimised by detecting invalid configurations earlier to avoid generating multiple invalid branches. Using this approach, all valid configurations were generated for systems with fewer than six sites and $m \le 4$ in under one minute. For eight sites at $m = 4$ , computational limitations prevented full enumeration, although the number of valid configurations is estimated to exceed $10^6$ .

n/m	1	2	3	4
2	6	18	66	258
4	18	90	546	3618
6	66	546	6840	94386
8	258	3618	94386	-

Table 1: Number of valid configurations for $n$ sites and different values of $m$

The generation of all world-lines enabled the computation of the Trotterization error which is shown in figure 6.

Metropolis algorithm

To perform a complete Monte Carlo simulation, the Metropolis algorithm was used and three different update schemes were implemented.

Local shifts

The first update approach is by shifting spin lines on a shaded plaquette.

Figure 2 (a): Configuration side

Figure 2 (b): Configuration crossed

The advantage of this technique is its ease of implementation as the energy and weight shifts are trivial to compute, and the move has a symmetric probability.

The major drawback of this update method is that it is not ergodic. Indeed, the number of up spins remains the same throughout the move, and same for the winding number. This restricts the simulation to a subspace of the problem.

Vertex updates

The second update method is inspired by the original papet and uses the mapping to the 6-vertex problem.

Figure 3 (a): Initial configuration

Figure 3 (b): Vertex representation

Figure 3 (c): Loop transformation

Figure 3 (d): Final configuration

Once the problem is mapped to the vertex representation, a random location is selected, and a random path is traced following the vertex orientations. This path necessarily returns to its starting point, forming a closed loop. The directions of the vertices encountered along the path are then inverted, producing another valid world-line configuration. Prior to the flip, the weight associated with the exchange along the loop is computed, as the move is subject to the Metropolis acceptance criterion. The transition probability is symmetric.

Any two configurations can be connected by flipping an appropriate set of loops, demonstrating ergodicity.

The configurations differ by a finite number of loops. Consider two configurations, $A$ and $B$ , in their six-vertex representations. For each white plaquette, either $A$ and $B$ share the same vertex configuration or they contain two arrows in opposite directions. In the latter case, exactly two arrows must be reversed, since each square possesses two outgoing and two incoming arrows. Because the arrows form continuous loops across all world-lines, the set of differing arrows between $A$ and $B$ corresponds to a collection of closed loops.
All loops are possible. Since paths are generated randomly and the visited plaquettes are chosen at random, any loop present in the configuration space can, in principle, be produced by the algorithm.

Loop updates

The last update method implemented is the one described in Assaad and Evertz.

It is first observed that each vertex configuration can be traversed in three distinct ways. Consequently, each plaquette is associated with three possible graphs, as illustrated in figure 4.

Figure 4: Correspondence between vertex drawings and graphs. Taken from Assaad and Evertz and adapted.

Weights are defined for each plaquette–graph combination as $W(S, G)$ , where $S \in \{1, 2, 3\}$ denotes the plaquette type and $G \in \{1, 2, 3, 4\}$ denotes the graph, such that the normalisation condition holds.

\begin{equation} W(S)=\sum_G W(S, G) \end{equation}

From figure 4, it follows that $W(1, 4) = W(2, 1) = W(3, 2) = 0$ .

These weights are used to define the probability associated with selecting a particular graph.

\begin{equation} \mathbb{P}\left(S \to (S, G)\right) = \frac{W(S, G)}{W(S)} \end{equation}

Once each plaquette has been mapped to a corresponding graph, multiple loops spanning the entire world-line configuration are obtained. A Union–Find data structure is employed to construct these loops, as illustrated in figure 5.

Figure 5: Partition of the word-line in loops. The green lines are the graphs. Note: the crossed lines represents graph 4 and not 3.

For each loop defined by the graphs, it is then determined whether to flip the spins. The probability of such a flip is computed accordingly:

\begin{equation} \mathbb{P}\left((S, G) \to (S', G)\right) = \frac{W(S', G)}{W(S, G) + W(S', G)} \end{equation}

To ensure symmetric transition probabilities in the Metropolis algorithm, it is required that for any $G$ and compatible $S$ and $S'$ , the symmetry condition holds:

\begin{align} W(S, G) = W(S', G) \end{align}

By combining equations 15 and 16, the following relation is obtained.

\begin{equation} \mathbb{P}\left((S, G) \to (S', G)\right) = \frac{1}{2} \end{equation}

Moreover the detailed balance is given by:

\begin{align*} W(S) \mathbb{P}\left(S\to S'\right) &= W(S)\sum_G \mathbb{P}\left(S\to (S, G)\right)\mathbb{P}\left((S, G) \to (S', G)\right) \\ &=\sum_G W(S)\frac{W(S, G)}{W(S)}\mathbb{P}\left((S, G) \to (S', G)\right) \\ &=\sum_G W(S', G)\mathbb{P}\left((S, G) \to (S', G)\right) \\ &=\sum_G \frac{W(S')}{W(S')} W(S', G)\mathbb{P}\left((S', G) \to (S, G)\right) \\ &=\sum_G W(S')\frac{W(S', G)}{W(S')} \mathbb{P}\left((S', G) \to (S, G)\right) \\ &= W(S')\sum_G \mathbb{P}\left(S'\to (S', G)\right)\mathbb{P}\left((S', G) \to (S, G)\right) \\ &= W(S') \mathbb{P}\left(S'\to S\right) \end{align*}

Thus, each loop has a probability of one half to be flipped.

The expressions for $W(S, G)$ are then derived. Following equation 13 and 16, the system of equations can be written as:

\begin{align*} W(1) &= W(1, 1) + W(1, 2) + W(1, 3) \\ W(2) &= W(1, 2) + W(2, 4) + W(1, 3) \\ W(3) &= W(1, 1) + W(2, 4) + W(1, 3) \end{align*}

The system contains four variables but only three independent equations. The parameter $W(1, 3) = f$ is selected as a free variable. After algebraic manipulation, the following expressions are obtained:

\begin{align*} W(1, 1) &= \frac{1}{2}\left(W(1) + W(3) - W(2) - f\right) \\ W(1, 2) &= \frac{1}{2}\left(W(1) + W(2) - W(3) - f\right) \\ W(2, 4) &= \frac{1}{2}\left(W(2) + W(3) - W(1) - f\right) \\ \end{align*}

To ensure that all graph weights remain positive, the resulting conditions on the weight values must be satisfied:

\begin{align*} W(1) &\leq W(2) + W(3) \\ W(2) &\leq W(1) + W(3) \\ W(3) &\leq W(1) + W(2) \end{align*}

For implementation simplicity, the parameter $f$ is set to zero, which restricts each plaquette to only two possible graphs.

The ergodicity of the method follows from the previous demonstration of ergodicity. Since any two configurations differ by a finite set of loops, and the present algorithm can generate and flip such loops, the method is therefore ergodic.

Discussion

Different update methods were compared by computing physical observables for small system sizes, where exact results are available for validation. These benchmarks were then extended to larger systems.

All Monte Carlo simulations were performed with identical parameters: $n_\text{cycle}=5000$ , $\text{length\_cycle}=2mn$ , $n_\text{rep}=10$ . Error bars represent the standard deviation across the $n_\text{rep}$ runs scaled down by $\sqrt{n_\text{rep}}$ .

The formulation from Assaad and Evertz was used to correct for Trotterization errors and compute observables:

\begin{equation} \langle O \rangle = \frac{\text{Tr} \left[ e^{-\beta H} O \right]}{\text{Tr} \left[ e^{-\beta H} \right]} = \frac{\sum_{w} \Omega(w) O(w)}{\sum_{w} \Omega(w)} \end{equation}

The largest system for which exact results were available was an $8$ -spin chain, simulated for several values of $m$ .

Energy

The energy was computed for the three update schemes. As discussed previously, the local shift method is not ergodic and therefore conserves the total spin, restricting the accessible configuration space.

Figure 6 (a): Energy as a function of $\Delta\tau$

Figure 6 (b): Error in log scale

The energy error for all methods is dominated by the Trotter approximation at large $\Delta\tau$ . However, for smaller values of $\Delta\tau$ ,the computed energies agree closely with the exact values, validating the model and suggesting the criterion $\Delta\tau \sim 0.1 J^{-1}$ for accurate calculations. \cref{fig:graphe_b2} indicates that the error scales as $\mathcal{O}(\Delta\tau^{1.83})$ , which is consistent with the expected $\mathcal{O}(\Delta\tau^{2})$ . The deviation observed for the vertex update is alter discussed.

The advantage of the numerical methods becomes clear for larger systems, such as 10 spin sites and $m = 6$ as can be seen in figure 7. The loop updates algorithm is able to calculate the energy for this larger system whereas the exact calculation was not possible.

Figure 7: Energy as a function of the number of sites for $m=6$

The model also allows computation of spin–spin correlation functions. Since each line represents a valid spin configuration, the correlation operator was averaged over all positions and imaginary-time slices:

\begin{equation} \left<S^z S^z_d\right> = \frac{1}{2mn} \sum_{j,i} S_{j,i}S_{j,i+d} \end{equation}

The exact correlation is computed as in equation 20 only on the first site (the value does not depend on the site).

\begin{equation} \left<S^z S^z_d\right> = \frac{\text{Tr}\left(e^{-\beta H}S^z_0S^z_d\right)}{\text{Tr}\left(e^{-\beta H}\right)} \end{equation}

The simulation accurately reproduced the correlations in both the ferromagnetic and antiferromagnetic regimes (figure 8).

Note: Other update schemes were not employed for this observable due to insufficient accuracy in their energy estimates.

Figure 8: Spin correlation computed using loop updates

The influence of the coupling $J_x$ on spin correlations was also examined (figure 9).

Figure 9: Spin correlation as a function of $J_x$ . _The graphs must be understood as follows : the right heat-map shows mean correlations values on the $n\\_\text{rep}$ runs and the left heat-map gives the uncertainty ( $\text{std}/\sqrt{n\\_\text{rep}}$ ) on each of the mean values of the right graph._

Efficiency of the different update methods

The efficiency of each scheme was evaluated using two metrics: computation time (figure 10) and acceptance rate (table 2).

Figure 10: Computation time for different update algorithms

m	2	4	6	8	10	12
Acceptance rate	46,6%	13,2%	5,9%	4,0%	3,3%	2.5%

Table 2: Acceptance rate for vertex update for $10^5$ tries

Although the loop update requires longer computation times, its acceptance rate remains close to $100\%$ , meaning that almost every proposed update results in a configuration change. However, for the vertex update the acceptance rate drops significantly as $m$ increases. Thus, for large $m$ , roughly $50\times$ more trials are required to accept an update. This can be seen in figure 6: for smaller $m$ , the accuracy of the vertex update decreases. This is probably due to an insufficient amount of accepted updates leading to inaccurate calculations. To conclude, the high acceptance rate of the loop update is essential for obtaining accurate results on large systems, justifying its longer computation time. This qualitative analysis could be verified by calculating the autocorrelation time which measures the necessary number of steps to obtain a decorrelated configuration. Efficiency could be defined as the inverse of the product of the autocorrelation and computation times.

In summary, the loop update proved to be the most accurate and efficient method implemented.

How this blog has been built

Sat, 20 Dec 2025 18:00:00 GMT

Throught my relentless scrolling of social medias I have heard of vibe codders. As I consider myself a coder I wanted to see if I could become also a vibe codder.

Previoulsy I used Claude Code in a 24h hackathon to quickly prototype a simple project. I was impressed by its power but since then I always worked on code where llms struggle to produce accurate code.

Everything changed when I got a free one year trial of Gemini Pro, I wanted to test how llms evolved since then and I thought that creating a blog was a great way of using this tool. Thus here we are !

The building with Gemini

My first attemp was to prompt to build a blog that read markdown files and produce a clean website using javascript. Gemini chose next.js but nothing worked and Gemini just kept messing with npm.

One rm -rf later

Off the bad start, I cleanned everything and reprompted him to use gatsby.js and specified to use the starter blog template, and ... it worked !

I then incremently added features to obtain this website ! It was pretty pleasant to just prompt Gemini and the real time changes happens !

Why Gatsby.js ?

It was a long time since I did not code website thus I used a tool that I used long ago, but it seems it was no longer really active. Nonetheless it works great !

I was able to add supports for code blocks, equations, images and even have a comment section using cusdis !

Aftertoughts & Conclusion

Building with Gemini was fun but also a bit creepy because now I have no idea of what part of the code does what.

Also javascript framework fatigue is real. I start doing physics (thus stop javascript) for 2 to 3 years and your favorite framework is now considered old !