<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Abhinav Anand on Medium]]></title>
        <description><![CDATA[Stories by Abhinav Anand on Medium]]></description>
        <link>https://medium.com/@abhinavanand1507?source=rss-972526d76de6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*7zxzR_2mxbqxlGlW.jpg</url>
            <title>Stories by Abhinav Anand on Medium</title>
            <link>https://medium.com/@abhinavanand1507?source=rss-972526d76de6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 21:34:20 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@abhinavanand1507/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Weight Entropy Dynamics During CNN Training: Preliminary Evidence of Entropy Stabilization During…]]></title>
            <link>https://medium.com/@abhinavanand1507/weight-entropy-dynamics-during-cnn-training-preliminary-evidence-of-entropy-stabilization-during-bef614d80f36?source=rss-972526d76de6------2</link>
            <guid isPermaLink="false">https://medium.com/p/bef614d80f36</guid>
            <category><![CDATA[entropy]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[backpropagation]]></category>
            <category><![CDATA[neural-networks]]></category>
            <dc:creator><![CDATA[Abhinav Anand]]></dc:creator>
            <pubDate>Thu, 12 Nov 2026 16:28:16 GMT</pubDate>
            <atom:updated>2026-11-12T16:28:16.134Z</atom:updated>
            <content:encoded><![CDATA[<h3>Weight Entropy Dynamics During CNN Training: Preliminary Evidence of Entropy Stabilization During Learning</h3><p><strong>A preliminary experimental study using MNIST and Fashion-MNIST</strong></p><p>refrence github repo : <a href="https://github.com/Abhinav1507/Deep-Learning/tree/main">https://github.com/Abhinav1507/Deep-Learning/tree/main</a></p><h3>Abstract</h3><p>Understanding how neural network parameters evolve during training is important for understanding optimization, learning, and generalisation. While training loss and accuracy are commonly used to monitor model performance, they do not directly describe how the internal distribution of network weights changes over time.</p><p>This preliminary study investigates the dynamics of weight entropy throughout a 200-epoch convolutional neural network (CNN) training experiment. Three entropy-based approaches were implemented: dynamic-range signed histogram entropy, fixed-bin signed histogram entropy, and magnitude-based Shannon entropy. The model was trained sequentially on MNIST and Fashion-MNIST, with dataset transitions at predefined training stages. Weight entropy, training accuracy, test accuracy, and loss measurements were recorded at every epoch.</p><p>The results show that magnitude-based weight entropy exhibits progressively smaller fluctuations during the initial MNIST training phase. In particular, the standard deviation of normalized magnitude entropy across the first ten epochs was approximately 0.00374, compared with approximately 0.00017 across epochs 81–90. Fixed-bin histogram entropy also showed a reduction in short-window variability, with its standard deviation decreasing from approximately 0.33701 to 0.01595 across the same windows.</p><p>These observations suggest that weight-entropy dynamics may provide additional information about the evolution of a neural network during training. In particular, entropy stabilization may be associated with a more stable learning regime. However, the findings are preliminary and do not establish a universal relationship between entropy stabilization and mathematical convergence. Further experiments involving multiple random seeds, independent training runs, gradient measurements, and parameter-update analysis are required.</p><p><strong>Keywords:</strong> Weight entropy, neural networks, CNN, training dynamics, Shannon entropy, convergence, parameter distribution, MNIST, Fashion-MNIST.</p><h3>1. Introduction</h3><p>Deep neural networks learn by repeatedly updating their parameters to minimize an objective function. During training, the model adjusts its weights in response to the difference between its predictions and the target labels. This optimization process gradually changes the internal representations of the network.</p><p>Training progress is commonly evaluated through measurements such as training loss, validation loss, and classification accuracy. These measurements describe how well a model performs, but they provide limited information about the evolution of its internal parameter distribution.</p><p>An interesting question is whether the distribution of neural network weights follows identifiable patterns throughout the training process.</p><p>In particular, does the entropy of a network’s weights change systematically as training progresses? Does entropy fluctuate substantially during early learning and become more stable as the network approaches convergence?</p><p>This study explores these questions by calculating several weight-entropy measurements at every epoch of CNN training.</p><p>The central idea investigated in this work is that <strong>weight entropy is not merely a static property of a trained neural network; it is a dynamic quantity whose evolution may provide information about the training process.</strong></p><p>The initial experimental observations suggest that entropy fluctuations become smaller during portions of training in which the model’s performance becomes relatively stable.</p><p>This leads to a preliminary research proposition:</p><blockquote><em>As a neural network progresses through training and approaches a stable optimization regime, the fluctuations in its weight-entropy measurements may decrease, resulting in increasingly stable entropy patterns.</em></blockquote><p>The present study investigates this proposition using a sequential training experiment involving MNIST and Fashion-MNIST.</p><h3>2. Theoretical Background</h3><h3>2.1 Shannon entropy and neural network weights</h3><p>Shannon entropy measures the uncertainty or dispersion of a probability distribution.</p><p>For a discrete probability distribution containing probabilities p₁, p₂, …, pₙ, Shannon entropy is defined as:</p><p><strong>H = − Σ pᵢ log₂(pᵢ)</strong></p><p>where the summation is taken over all N elements.</p><p>In this study, entropy is applied to neural network parameters to describe different aspects of their distribution.</p><p>It is important to recognize that entropy depends on how the weights are represented and how the probability distribution is constructed. Therefore, different entropy definitions can reveal different characteristics of the same network.</p><h3>2.2 Dynamic-range histogram entropy</h3><p>The first measurement calculates entropy from the signed distribution of weights.</p><p>The weights are flattened, and a histogram is constructed using 100 bins. The histogram range is determined using the minimum and maximum weight values at the current epoch.</p><p>If qᵦ represents the proportion of weights falling into histogram bin b, then:</p><p><strong>H_dynamic = − Σ qᵦ log₂(qᵦ)</strong></p><p>where:</p><ul><li>B = 100 histogram bins.</li><li>qᵦ = proportion of weights in bin b.</li></ul><p>This measurement describes how signed weight values are distributed across the current weight range.</p><p>Because the histogram range is recalculated at each epoch, changes in this entropy should not be interpreted as changes measured on an identical absolute weight scale.</p><h3>2.3 Fixed-bin histogram entropy</h3><p>The second measurement uses fixed histogram boundaries.</p><p>The weight values are clipped to the interval [−1, 1], and the same 100 histogram bins are used throughout training.</p><p>The resulting entropy is:</p><p><strong>H_fixed = − Σ qᵦ log₂(qᵦ)</strong></p><p>Unlike the dynamic-range method, this approach provides a consistent histogram reference across epochs.</p><p>However, values outside the interval are assigned to the boundary bins. Consequently, the result should be interpreted as clipped fixed-bin entropy, and the proportion of weights outside the interval should be monitored in future experiments.</p><h3>2.4 Magnitude-based weight entropy</h3><p>The third measurement focuses on the distribution of absolute weight magnitudes.</p><p>For weights w₁, w₂, …, wₙ, define the relative magnitude of each weight as:</p><p><strong>pᵢ = |wᵢ| / Σⱼ₌₁ᴺ |wⱼ|</strong></p><p>The magnitude entropy is then:</p><p><strong>H_magnitude = − Σᵢ₌₁ᴺ pᵢ log₂(pᵢ)</strong></p><p>To facilitate comparison across parameter collections of different sizes, normalized magnitude entropy is calculated:</p><p><strong>H_normalized = H_magnitude / log₂(N)</strong></p><p>A higher value indicates a relatively more evenly distributed magnitude contribution, while a lower value indicates greater concentration of total absolute weight magnitude among a smaller subset of parameters.</p><p>This measurement does not directly establish the functional importance of individual weights. Rather, it describes how the total absolute magnitude is distributed across parameters.</p><h3>2.5 Entropy stabilization</h3><p>The entropy value at epoch t can be represented as Hₜ.</p><p>However, the entropy value alone does not describe its temporal stability.</p><p>To investigate fluctuations, consider the standard deviation of entropy over a window of W epochs:</p><p><strong>F_H(t) = √[(1/W) × Σᵢ₌ₜ₋W₊₁ᵗ (Hᵢ − H̄ₜ)²]</strong></p><p>where:</p><ul><li>F_H(t) = entropy fluctuation at epoch t.</li><li>W = number of epochs in the rolling window.</li><li>Hᵢ = entropy at epoch i.</li><li>H̄ₜ = average entropy within the rolling window.</li></ul><p>A reduction in F_H(t) indicates that entropy values are becoming less variable over that window.</p><p>This distinction is important:</p><ul><li>Entropy level describes the current distribution of weights.</li><li>Entropy fluctuation describes how much that distribution changes over time.</li></ul><p>A neural network may have relatively high entropy with very little fluctuation, or low entropy with substantial fluctuation.</p><p>Therefore, this study considers entropy stabilization separately from entropy reduction.</p><h3>3. Experimental Methodology</h3><h3>3.1 Model and training configuration</h3><p>A convolutional neural network was implemented using PyTorch.</p><p>The experiment used the same CNN architecture for MNIST and Fashion-MNIST, both of which contain grayscale images and ten classification classes.</p><p>The main training configuration included:</p><p>ParameterConfigurationModelCNNFrameworkPyTorchOptimizerAdamLearning rate0.001Batch size64Loss functionCross-entropy lossHistogram bins100Fixed histogram range[−1, 1]Random seed42Total epochs200</p><p>Weight entropy was calculated after every training epoch.</p><p>The measurements were collected for individual weight-bearing layers and for all selected weights pooled together. Bias parameters were not included in the weight-entropy calculations.</p><h3>3.2 Sequential dataset training schedule</h3><p>The experiment consisted of four consecutive training phases:</p><p>EpochsDatasetNumber of epochs1–90MNIST9091–180Fashion-MNIST90181–190MNIST10191–200Fashion-MNIST10</p><p>An important feature of this experiment is that training was resumed using saved model checkpoints.</p><p>The checkpoint included the model parameters and Adam optimizer state. Therefore, the experiment represents sequential adaptation of the same model rather than four independently initialized training experiments.</p><p>This setup makes it possible to observe entropy dynamics when the training dataset changes, while retaining the model’s previously learned parameters.</p><h3>3.3 Data collection</h3><p>Three CSV files were generated during training:</p><ol><li><strong>Weight entropy history:</strong> Records entropy measurements for individual layers and the combined network at every epoch.</li><li><strong>Accuracy history:</strong> Records training accuracy, test accuracy, accuracy gap, training loss, and test loss.</li><li><strong>Training log:</strong> Records the principal training and evaluation measurements alongside combined entropy values.</li></ol><p>Together, these files provide an epoch-wise record of parameter-distribution measurements and model performance throughout the experiment.</p><h3>4. Experimental Results</h3><h3>4.1 Training performance</h3><p>The experiment achieved the following average performance across its four phases:</p><p>Training phaseMean training accuracyMean test accuracyMNIST, epochs 1–9099.76%99.28%Fashion-MNIST, epochs 91–18097.99%91.67%MNIST, epochs 181–19099.65%99.23%Fashion-MNIST, epochs 191–20099.21%91.84%</p><p>The results indicate that the network achieved high classification accuracy on both datasets, although Fashion-MNIST presented a more difficult classification problem.</p><p>During the first MNIST phase, training and test accuracy remained high. During Fashion-MNIST training, the difference between training and test accuracy was larger.</p><p>These differences provide useful context when interpreting entropy changes across the sequential training phases.</p><h3>4.2 Entropy measurements across training phases</h3><p>The mean combined entropy values were:</p><p>Training phaseDynamic-range entropyFixed-bin entropyMagnitude entropy (bits)Normalized magnitude entropyMNIST, 1–903.84104.121518.30940.9516Fashion-MNIST, 91–1804.04534.946218.20710.9463MNIST, 181–1903.98045.129418.19220.9455Fashion-MNIST, 191–2003.98985.158518.19330.9456</p><p>Several observations emerge.</p><p>First, the dynamic-range histogram entropy and fixed-bin histogram entropy show different patterns from magnitude entropy.</p><p>Second, magnitude entropy generally decreases over the longer training sequence, indicating an increasingly concentrated distribution of absolute parameter magnitudes under this particular measurement.</p><p>Third, entropy behaviour changes across dataset transitions. This suggests that different training phases may influence the distribution of network parameters differently.</p><p>These measurements demonstrate why multiple entropy definitions are useful: a single entropy measurement does not capture every aspect of the weight distribution.</p><h3>4.3 Entropy fluctuation during the initial MNIST training phase</h3><p>The most interesting observation concerns the reduction in short-window entropy variability during the first 90 epochs.</p><p>For this analysis, the combined entropy values were divided into consecutive windows of ten epochs. The standard deviation was calculated within each window.</p><p>The following results compare the first ten epochs with the final ten epochs of the initial MNIST phase.</p><p>Entropy measurementEpochs 1–10 standard deviationEpochs 81–90 standard deviationDynamic-range histogram entropy0.246870.00925Fixed-bin histogram entropy0.337010.01595Magnitude entropy0.072010.00331Normalized magnitude entropy0.003740.00017</p><p>The reduction is substantial across all four measurements.</p><p>For normalized magnitude entropy, the standard deviation decreased from approximately 0.00374 to 0.00017.</p><p>This represents a reduction of approximately 95% in the within-window standard deviation.</p><p>Similarly, the fixed-bin histogram entropy standard deviation decreased from approximately 0.33701 to 0.01595.</p><p>The dynamic-range histogram entropy also became less variable during this phase.</p><p>These observations provide preliminary evidence supporting the idea that weight-entropy fluctuations become smaller as training progresses through a stable learning phase.</p><p>Importantly, the fixed-bin entropy and magnitude entropy continued to show gradual directional changes. Thus, stabilization of fluctuations does not necessarily mean that the entropy value itself becomes completely constant.</p><h3>4.4 Entropy behaviour during later Fashion-MNIST training</h3><p>During the Fashion-MNIST phase from epochs 91–180, the combined normalized magnitude entropy showed relatively small variability.</p><p>Its mean was approximately 0.94628, with a standard deviation of approximately 0.00082.</p><p>During the final ten epochs of this phase, epochs 171–180, the standard deviation decreased further to approximately 0.00006.</p><p>The fixed-bin histogram entropy also showed a reduction in short-window variability:</p><ul><li>Epochs 91–100: standard deviation approximately 0.02995.</li><li>Epochs 171–180: standard deviation approximately 0.00663.</li></ul><p>These observations are consistent with continued entropy stabilization during later training.</p><p>However, the dynamic-range histogram entropy behaved differently and exhibited larger fluctuations in some later windows.</p><p>This difference reinforces the importance of distinguishing between entropy definitions when drawing conclusions about training dynamics.</p><h3>4.5 Entropy behaviour during dataset transitions</h3><p>The sequential experiment also provides an opportunity to observe entropy changes when the training dataset changes.</p><p>At the transition from epoch 190 to epoch 191, the combined entropy measurements changed as follows:</p><p>MeasurementEpoch 190Epoch 191Dynamic-range entropy3.896514.10040Fixed-bin entropy5.138975.15001Magnitude entropy18.1939418.19442Normalized magnitude entropy0.945590.94562</p><p>The dynamic-range entropy showed a more noticeable change than the other measurements.</p><p>This may reflect a change in the weight distribution associated with the new training phase. However, because the experiment does not isolate dataset effects from optimizer dynamics and other training factors, this observation should be treated as descriptive rather than causal.</p><h3>5. Discussion</h3><h3>5.1 Entropy stabilization as a potential indicator of training stability</h3><p>The central observation of this experiment is that entropy fluctuations decreased substantially during the initial MNIST training phase.</p><p>The reduction was visible in all four combined entropy measurements, although the magnitude differed between methods.</p><p>The same general tendency was also observed in the fixed-bin and magnitude-based entropy measurements during the later Fashion-MNIST training phase.</p><p>This suggests that entropy dynamics may contain information about changes in the internal parameter distribution as a network learns.</p><p>One possible interpretation is that early training involves more substantial redistribution of weight values, while later training produces smaller changes in the measured distribution.</p><p>As the model approaches a more stable learning regime, the statistical structure of its weight distribution may become increasingly stable.</p><p>This leads to the possibility that weight-entropy fluctuation could be investigated as an additional indicator of optimization stability.</p><p>However, entropy stabilization should not be equated with mathematical convergence. Entropy is a many-to-one summary of a parameter distribution: different weight configurations can have identical entropy values. A network’s weights can also continue changing while its entropy remains nearly constant.</p><p>Direct measurements of gradient norms, parameter-update magnitudes, and loss trajectories are needed to establish whether entropy stabilization coincides with optimization convergence.</p><h3>5.2 Entropy stabilization does not necessarily mean entropy reduction</h3><p>An important distinction emerges from the results.</p><p>Magnitude entropy generally decreased across the longer training sequence, while fixed-bin histogram entropy generally increased.</p><p>Nevertheless, both measurements showed reduced short-window fluctuations during parts of training.</p><p>This indicates that entropy stabilization and entropy reduction are different phenomena.</p><p>A network may reach a more stable weight-distribution regime without its entropy necessarily decreasing.</p><p>Consequently, the proposed research framework should investigate two separate properties:</p><ol><li>The direction and magnitude of entropy change.</li><li>The temporal stability of entropy.</li></ol><p>This distinction may be important for developing entropy-based descriptions of neural network training.</p><h3>5.3 Potential relationship with parameter specialization</h3><p>The decrease in magnitude entropy suggests that the absolute magnitudes of the network’s parameters became more concentrated under the probability distribution used in this study.</p><p>One possible interpretation is that the relative contribution of a subset of parameters to total absolute weight magnitude increased.</p><p>This could be related to parameter specialization, where particular parameters or groups of parameters become more prominent in the learned representation.</p><p>However, magnitude concentration alone does not establish that these parameters have greater functional influence on predictions.</p><p>Future experiments should investigate whether perturbing or removing the largest-magnitude parameters causes a greater change in model performance than perturbing an equivalent number of randomly selected parameters.</p><h3>5.4 Potential relationship with generalisation</h3><p>The experiment also shows that high training accuracy does not necessarily imply equally high test accuracy.</p><p>Fashion-MNIST exhibited a larger average training-test accuracy gap than MNIST.</p><p>This provides an opportunity to investigate whether entropy dynamics are associated with changes in generalisation behaviour.</p><p>However, the current results do not establish that decreasing entropy causes poorer generalisation or overfitting.</p><p>A decrease in magnitude entropy could represent useful parameter organization, excessive specialization, or another change in parameter scale and distribution.</p><p>Distinguishing these possibilities requires controlled generalisation and robustness experiments.</p><h3>6. Limitations</h3><p>Several limitations should be considered when interpreting these findings.</p><p><strong>First, the experiment uses a single sequential training run.</strong> The results have not yet been replicated across multiple random initializations.</p><p><strong>Second, the experiment does not directly measure optimization convergence.</strong> Gradient norms and parameter-update magnitudes were not included in the recorded CSV files. Therefore, entropy stabilization is evidence of stability in the measured entropy statistics, not proof of convergence.</p><p><strong>Third, the training dataset changes during the experiment.</strong> The model retains its weights and optimizer state when moving between MNIST and Fashion-MNIST. Consequently, the later phases are not independent experiments.</p><p><strong>Fourth, the entropy methods measure different properties.</strong> Dynamic-range histogram entropy uses an epoch-specific range, fixed-bin entropy clips weights to a fixed interval, and magnitude entropy measures the relative distribution of absolute parameter values.</p><p><strong>Fifth, the fixed-bin measurement does not record the proportion of weights outside the clipping interval.</strong> This should be included in future experiments.</p><p><strong>Sixth, training accuracy is calculated during training mode in the supplied implementation, whereas test accuracy is calculated in evaluation mode.</strong> If the CNN contains BatchNorm or Dropout, this can affect direct comparisons between the two measurements. A separate evaluation-mode calculation of training accuracy would provide a more controlled comparison.</p><p>Finally, the current experiment evaluates the dataset used in each training phase. It does not evaluate both datasets at every phase boundary, so it cannot establish whether sequential training caused forgetting.</p><h3>7. Future Research Directions</h3><p>The findings motivate several follow-up investigations.</p><h3>7.1 Entropy stabilization and optimization convergence</h3><p>Future experiments should record gradient norms and relative parameter-update magnitudes at every epoch.</p><p>The objective would be to determine whether reduced entropy fluctuation occurs alongside reduced parameter-update magnitudes and stabilization of training loss.</p><h3>7.2 Reproducibility across random seeds</h3><p>The experiment should be repeated with multiple random seeds and independent training runs.</p><p>This would help determine whether entropy stabilization is a reproducible characteristic of CNN training rather than an observation specific to one initialization.</p><h3>7.3 Layer-wise entropy dynamics</h3><p>Entropy should be investigated separately for convolutional and classifier layers.</p><p>Different layers may exhibit different entropy trajectories because they perform different computational roles.</p><h3>7.4 Parameter concentration and sensitivity</h3><p>Future experiments should compare the effects of removing or perturbing dominant-magnitude weights against random parameter perturbations.</p><p>This could establish whether decreasing magnitude entropy is associated with greater dependence on a smaller subset of parameters.</p><h3>7.5 Generalisation and overfitting</h3><p>The relationship between entropy dynamics and generalisation should be studied using independent training and test evaluations.</p><p>Particular attention should be given to whether entropy reduction is associated with useful specialization or with excessive dependence on dataset-specific features.</p><h3>7.6 Generalisation across architectures</h3><p>After establishing reproducible findings in CNNs, similar experiments could be conducted using ResNet architectures and other neural network models.</p><p>This would help determine whether the observed entropy patterns are specific to the architecture or represent broader characteristics of neural network optimization.</p><h3>8. Conclusion</h3><p>This preliminary study investigated the evolution of weight entropy during a 200-epoch CNN training experiment involving MNIST and Fashion-MNIST.</p><p>Three entropy-based approaches were implemented to describe different aspects of the network’s weight distribution.</p><p>The most notable finding was the substantial reduction in entropy fluctuations during the initial MNIST training phase. The standard deviation of normalized magnitude entropy across ten-epoch windows decreased from approximately 0.00374 during epochs 1–10 to approximately 0.00017 during epochs 81–90. Fixed-bin histogram entropy showed a similar reduction in variability.</p><p>Later Fashion-MNIST training also exhibited relatively small fluctuations in magnitude entropy and fixed-bin histogram entropy.</p><p>These findings support further investigation of the hypothesis that <strong>weight-entropy fluctuations decrease as a neural network progresses toward a more stable training regime.</strong></p><p>The experiment also highlights an important distinction between entropy stabilization and entropy reduction. While magnitude entropy generally decreased, fixed-bin histogram entropy increased across substantial portions of training. Thus, the temporal stability of entropy may be a separate and potentially useful property from the absolute entropy level.</p><p>The broader research direction is to investigate whether weight-entropy dynamics can provide additional information about the training life cycle of neural networks, including optimization stability, parameter concentration, specialization, and generalisation.</p><p>At this stage, the results should be interpreted as preliminary experimental evidence rather than a universal theoretical law. Nevertheless, they provide a foundation for a more systematic investigation of weight entropy as a dynamic characteristic of neural network learning.</p><p><strong>The central proposition emerging from this experiment is that the evolution and stabilization of weight entropy may offer an additional perspective on the internal dynamics of neural network training, beyond conventional performance measurements such as loss and accuracy.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bef614d80f36" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A Coordination-Based Discrete Lattice Model for Relativistic Motion]]></title>
            <link>https://medium.com/@abhinavanand1507/a-coordination-based-discrete-lattice-model-for-relativistic-motion-553d2ec0854a?source=rss-972526d76de6------2</link>
            <guid isPermaLink="false">https://medium.com/p/553d2ec0854a</guid>
            <category><![CDATA[general-relativity]]></category>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[spacetime]]></category>
            <category><![CDATA[special-relativity]]></category>
            <category><![CDATA[relativity]]></category>
            <dc:creator><![CDATA[Abhinav Anand]]></dc:creator>
            <pubDate>Sun, 12 Apr 2026 16:09:13 GMT</pubDate>
            <atom:updated>2026-04-12T16:09:13.463Z</atom:updated>
            <content:encoded><![CDATA[<h3>A Coordination-Based Discrete Lattice Model for Relativistic Motion</h3><h3>Abstract</h3><p>We present a discrete lattice model of spacetime in which particle mass is associated with the number of lattice coordinates it occupies. Motion arises from coordinated updates across these coordinates, while energy is interpreted as the ability to achieve such coordination. By identifying rest mass as (m = N mₗ), where (N) is the number of occupied lattice sites and (mₗ) is a fundamental mass unit, we reinterpret relativistic energy as a sum over discrete degrees of freedom. The model reproduces key features of Special Relativity, including the existence of a maximum speed © and the requirement of infinite energy for massive particles to reach it.</p><h3>1. Introduction</h3><p>In Special Relativity, the speed of light © is an invariant upper bound for any particle with nonzero rest mass. As velocity approaches ©, the required energy diverges.</p><p>Motivated by the possibility that spacetime may possess a discrete microscopic structure, we construct a lattice-based model in which relativistic behavior emerges from coordination constraints among discrete degrees of freedom.</p><h3>2. Discrete Lattice Structure</h3><p>Spacetime is modeled as a lattice with coordinates:</p><p>x = n Δx<br>t = m Δt</p><p>where n and m are integers, and Δx and Δt are fundamental spacings.</p><p>The fundamental speed is defined as:</p><p>c = Δx / Δt</p><p>Each lattice point represents a spacetime event.</p><h3>3. Particle Representation</h3><p>A particle is defined as a set of lattice coordinates:</p><p>S(t) = {x₁(t), x₂(t), … , x_N(t)}</p><p>where N is the number of occupied lattice sites.</p><p>We define the rest mass as:</p><p>m = N mₗ</p><p>where mₗ is the mass associated with a single lattice coordinate.</p><p>Since rest mass is invariant under changes of reference frame, both N and mₗ are taken to be invariant quantities.</p><h3>4. Discrete Evolution Rule</h3><p>At each time step Δt, each coordinate evolves as:</p><p>xᵢ(t + Δt) = xᵢ(t) + δxᵢ</p><p>where:</p><p>δxᵢ ∈ {−Δx, 0, +Δx}</p><p>A successful propagation step requires:</p><p>δx₁ = δx₂ = … = δx_N</p><p>Thus, motion occurs only when all coordinates update coherently.</p><h3>5. Velocity and Coordination</h3><p>If a coordinated update occurs:</p><p>Δx = δx</p><p>Velocity is defined as:</p><p>v = Δx / Δt</p><p>Thus, instantaneous velocity takes values:</p><p>v ∈ {−c, 0, +c}</p><p>Over many steps, the average velocity is:</p><p>v = p c</p><p>where p represents the probability of successful coordinated motion.</p><h3>6. Relativistic Energy</h3><p>We adopt the standard relativistic expression:</p><p>E = γ m c²</p><p>with:</p><p>γ = 1 / √(1 − v² / c²)</p><p>Substituting m = N mₗ:</p><p>E = γ N mₗ c²</p><p>Thus, total energy is interpreted as the sum of contributions from N lattice coordinates.</p><h3>7. Energy–Momentum Relation</h3><p>The relativistic relation is:</p><p>E² = pₘ² c² + m² c⁴</p><p>Substituting m = N mₗ:</p><p>E² = pₘ² c² + N² mₗ² c⁴</p><p>Velocity is given by:</p><p>v = (pₘ c²) / E</p><p>which leads to:</p><p>v = c √(1 − (N² mₗ² c⁴) / E²)</p><h3>8. Emergence of the Speed Limit</h3><p>As velocity approaches c:</p><p>(N² mₗ² c⁴) / E² → 0</p><p>This requires:</p><p>E → ∞</p><p>Thus, a particle with nonzero N (mass) requires infinite energy to reach the speed of light.</p><h3>9. Time Dilation</h3><p>Relativistic time dilation is:</p><p>dτ = dt √(1 − v² / c²)</p><p>Substituting the expression for γ:</p><p>dτ = dt / γ</p><p>Using the energy relation:</p><p>γ = E / (N mₗ c²)</p><p>Thus:</p><p>dτ = dt (N mₗ c² / E)</p><p>This suggests that proper time corresponds to internal lattice evolution, which slows as energy increases.</p><h3>10. Interpretation</h3><p>In this framework:</p><ul><li>Mass corresponds to the number of lattice coordinates</li><li>Motion requires coordinated updates across coordinates</li><li>Energy represents the ability to achieve coordination</li><li>The speed limit emerges from the requirement of perfect coordination</li></ul><p>Massless particles (N = 0 effective limit) propagate at speed c, while massive particles (N &gt; 0) cannot reach c.</p><h3>11. Conclusion</h3><p>We have developed a discrete lattice model in which relativistic behavior is interpreted in terms of coordination among discrete degrees of freedom. By identifying rest mass with the number of lattice coordinates, the relativistic energy expression is reinterpreted as a sum over fundamental units.</p><p>The model reproduces key features of Special Relativity while providing a possible microscopic picture of spacetime structure. Further work is required to derive relativistic relations directly from lattice dynamics and to explore potential observable consequences.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=553d2ec0854a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Dynamics of Photon and Spacetime]]></title>
            <link>https://medium.com/@abhinavanand1507/dynamics-of-photon-and-spacetime-1c19206e1f71?source=rss-972526d76de6------2</link>
            <guid isPermaLink="false">https://medium.com/p/1c19206e1f71</guid>
            <dc:creator><![CDATA[Abhinav Anand]]></dc:creator>
            <pubDate>Mon, 27 Sep 2021 07:38:31 GMT</pubDate>
            <atom:updated>2021-09-27T07:38:31.700Z</atom:updated>
            <content:encoded><![CDATA[<p>ABSTRACT: Photon does show some important quantum mechanical properties and it’s investigation has always been important throughout the history of quantum mechanics. Here I am trying to understand the nature of a photon’s speed ‘c’ and its relation with gravity. Here I am using the concept of both GR and QM. Goal of this paper is also to understand the interaction of spacetime with energy.</p><p>INTRODUCTION: We know the value of speed of light(c) &amp; Relativity(General and Special). Here I am trying to understand the reason for it being constant in Relativity theory.</p><p>ASSUMPTION:</p><ol><li>Here I am considering a quantized version of spacetime with metric tensor (Sij) and Volume (Vs ) based on metric. We can visualize it as 4-dimension hole.</li><li>2)Quantized spacetime is the building block of curvature/geodesic through which gravity works.</li><li>3)In presence of Energy, spacetime gets compressed to create curvature for gravity</li></ol><p>POSTULATES:</p><p>1)In the presence of Energy, spacetime gets compressed to create curvature for gravity. Because of compression, volume of quantized space time decrease</p><p>2) space occupied by energy(Volume of energy Vt) must be less than Vs for speed to be ‘c’. It is an energy limit till which particle can have speed c.</p><p>As we know,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/180/1*XHFI5z9r81CNCNDQws0lGQ.png" /></figure><p>and for photon E=hc/λ . We do know that because of GR or we can also think of it as quantized spacetime compression, Energy of the photon would increase and it would become heavier.</p><p>If we consider the system of only one photon, we know a sudden increase in energy would not be possible because of conservation of energy. Now Because of quantized spacetime compression , the energy limit ‘Ts’(till which particle can have speed c) would decrease. Difference between current E and current energy limit Ts would emerge as current internal energy(current mass ‘mo’). This internal energy may break the particle (heavy photon) in more than one particle. One with speed c and other with mass .</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/555/1*5k1mDexNdXHSSM8-2OOC6A.png" /></figure><p>Conclusion: Considering the postulates which hold with GR and QM , we can accept space / spacetime is quantized as discrete and space quantum does have volume. space quantum is hole but higher dimension hole. It also gives light it’s constant speed.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1c19206e1f71" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>