Can Et Jaynes Probability Theory Explain Bayesian Model Selection?

I've been nerding out over Jaynes for years and his take feels like a breath of fresh air when frequentist methods get too ritualistic. Jaynes treats probability as an extension of logic — a way to quantify rational belief given the information you actually have — rather than merely long-run frequencies. He leans heavily on Cox's theorem to justify the algebra of probability and then uses the principle of maximum entropy to set priors in a principled way when you lack full information. That means you don't pick priors by gut or convenience; you encode symmetry and constraints, and let entropy give you the least-biased distribution consistent with those constraints. By contrast, the frequentist mindset defines probability as a limit of relative frequencies in repeated experiments, so parameters are fixed and data are random. Frequentist tools like p-values and confidence intervals are evaluated by their long-run behavior under hypothetical repetitions. Jaynes criticizes many standard procedures for violating the likelihood principle and being sensitive to stopping rules — things that, from his perspective, shouldn't change your inference about a parameter once you've seen the data. Practically that shows up in how you interpret intervals: a credible interval gives the probability the parameter lies in a range, while a confidence interval guarantees coverage across repetitions, which feels less directly informative to me. I like that Jaynes connects inference to decision-making and prediction: you get predictive distributions, can incorporate real prior knowledge, and often get more intuitive answers in small-data settings. If I had one tip, it's to try a maximum-entropy prior on a toy problem and compare posterior predictions to frequentist estimates — it usually opens your eyes.

If I had to boil Jaynes down to a handful of guiding lights, they'd be: probability as extended logic, maximum entropy as the least biased assignment given constraints, and symmetry/invariance for choosing priors. I love how Jaynes treats probabilities not as long-run frequencies but as degrees of plausibility — numbers that obey rational rules (think Cox's desiderata) so different lines of reasoning give consistent results. He pushes the maximum entropy principle hard: when all you know are some constraints (like averages), choose the distribution that maximizes Shannon entropy subject to those constraints. That way you don't smuggle in extra assumptions. He also insists priors should reflect symmetry and transformation groups — use the problem's invariances to pick noninformative priors rather than an ill-defined “ignorance.” Finally, and this is the practical kicker, update with Bayes' rule when you get data, and always be explicit about what information you're conditioning on. I keep a copy of 'Probability Theory: The Logic of Science' on my shelf and treat it like a toolkit: logic for setting up plausibilities, MaxEnt for turning constraints into distributions, and invariance arguments for fair priors.

Okay, dive in with me: if you only take a few chapters from 'Probability Theory: The Logic of Science', I’d grab the ones that build the whole way you think about uncertainty. Start with Jaynes’s foundational material — the chapters that explain probability as extended logic and derive the product and sum rules. Those are the philosophical and mathematical seeds that make the rest of the book click; without them, Bayes' theorem and conditionals feel like magic tricks instead of tools. After that, read the section on prior probabilities and transformation groups: Jaynes’s treatment of invariance and how to pick noninformative priors is pure gold, and it changes how you set up problems. Then move to the parts on the method of maximum entropy and on parameter estimation/approximation methods. Maximum entropy is the cleanest bridge between information theory and inference, and the estimation chapters show you how to actually compute credible intervals and compare models. If you like case studies, skim the applied chapters (spectral analysis, measurement errors) later; they show the ideas in action and are surprisingly practical. Personally, I flip between the core theory and the examples — theory to understand, examples to remember how to use it.

I get a little giddy whenever Jaynes comes up because his way of thinking actually makes prior selection feel like crafting a story from what you truly know, not just picking a default. In my copy of 'Probability Theory: The Logic of Science' I underline whole paragraphs that insist priors should reflect symmetries, invariances, and the constraints of real knowledge. Practically that means I start by writing down the facts I have — what units are natural, what quantities are invariant if I relabel my data, and what measurable constraints (like a known average or range) exist. From there I often use the maximum entropy principle to turn those constraints into a prior: if I only know a mean and a range, MaxEnt gives the least-committal distribution that honors them. If there's a natural symmetry — like a location parameter that shifts without changing the physics — I use uniform priors on that parameter; for scale parameters I look for priors invariant under scaling. I also do sensitivity checks: try a Jeffreys prior, a MaxEnt prior, and a weakly informative hierarchical prior, then compare posterior predictions. Jaynes’ framework is a mindset as much as a toolbox: encode knowledge transparently, respect invariance, and test how much your conclusions hinge on those modeling choices.

When I flip through problems inspired by Jaynes, the classics always pop up: biased coin estimation, urn problems, dice symmetry, and the ever-delicious applications of maximum entropy. A typical exercise will have you infer the bias of a coin after N tosses using a Beta prior, or derive the posterior predictive for the next toss — that little sequence of Beta-Binomial calculations is like comfort food. Jaynes also loves urn problems and variations on Bertrand's paradox, where you wrestle with what the principle of indifference really means and how choices of parameterization change probabilities. He then stretches those ideas into physics and information theory: deriving the Gaussian, exponential, and Poisson distributions from maximum-entropy constraints, or getting the canonical ensemble by maximizing entropy with an energy constraint. I've used those exercises to explain how statistical mechanics and Bayesian inference are cousins, and to show friends why the 'right' prior sometimes comes from symmetry or from maximum entropy. Throw in Monty Hall style puzzles, Laplace’s rule of succession, and simple sensor-noise inference examples and you’ve covered most of the recurring motifs — problems that are conceptually elegant but also great for coding quick Monte Carlo checks.

If Jaynes' 'Probability Theory: The Logic of Science' lit a fire for you, I found the natural next steps split into three flavors: conceptual, applied, and rigorous math. On the conceptual/Bayesian side I keep going back to 'Bayesian Data Analysis' by Gelman et al. — it’s expansive, honest about practical pitfalls, and full of real examples. For a warm, conversational bridge between intuition and practice, 'Statistical Rethinking' by Richard McElreath rewired the way I build models: his code-first, example-driven approach makes Bayesian ideas stick. If you want a very hands-on, tutorial-style companion, John Kruschke’s 'Doing Bayesian Data Analysis' is delightful. For computational and machine-learning perspectives, Kevin P. Murphy’s 'Machine Learning: a Probabilistic Perspective' and Bishop’s 'Pattern Recognition and Machine Learning' show how probabilistic thinking powers algorithms. For foundational probability with measure-theoretic rigor, 'Foundations of Modern Probability' by Olav Kallenberg is brutal but rewarding, and Rick Durrett’s 'Probability: Theory and Examples' balances clarity with depth. I usually alternate between these books depending on whether I need intuition, code, or proofs.

Okay, quick and friendly: if you want a legal download of E. T. Jaynes' famous book, look first at the publisher. Cambridge University Press sells electronic versions of 'Probability Theory: The Logic of Science' — that's the most straightforward, aboveboard way to get a PDF or an ebook copy. If you have access through a university, your library might already subscribe to Cambridge e-books, so you could read or download it via your institution. Another legit route is major ebook vendors: Google Play Books and Amazon (Kindle) often carry the title. Those aren’t always PDFs, but they’re licensed ebooks you can buy immediately. If buying isn’t an option, try your local or university library: WorldCat can show nearby physical copies and many libraries participate in interlibrary loan if they don’t own it. Finally, check Open Library/Internet Archive for a borrowable digital copy — they lend legally under controlled digital lending. If you’re unsure whether a PDF you find online is legal, follow the publisher’s page or contact them directly; I’ve done that once and they were helpful. Happy reading — it’s a dense, brilliant book, so get a comfy chair and good coffee.

What keeps Jaynes on reading lists and citation trails decades after his papers? For me it's the mix of clear philosophy, practical tools, and a kind of intellectual stubbornness that refuses to accept sloppy thinking. When I first dug into 'Probability Theory: The Logic of Science' I was struck by how Jaynes treats probability as extended logic — not merely frequencies or mystical priors, but a coherent calculus for reasoning under uncertainty. That reframing still matters: it gives people permission to use probability where they actually need to make decisions. Beyond philosophy, his use of Cox's axioms and the maximum entropy principle gives concrete methods. Maximum entropy is a wonderfully pragmatic rule: encode what you know, and otherwise stay maximally noncommittal. I find that translates directly to model-building, whether I'm sketching a Bayesian prior or cleaning up an ill-posed inference. Jaynes also connects probability to information theory and statistical mechanics in ways that appeal to both physicists and data people, so his work lives at multiple crossroads. Finally, Jaynes writes like he’s hashing things out with a friend — opinionated, rigorous, and sometimes cranky — which makes the material feel alive. People still cite him because his perspective helps them ask better questions and build cleaner, more honest models. For me, that’s why his voice keeps showing up in citation lists and lunchtime debates.