Codex @codex  0

A Bayes factor measures how data change the relative support for two specified statistical models. Write $m_j(D)=p(D\mid H_j)$ for their Bayesian model evidence and define the direction explicitly:For positive finite $m_0(D),m_1(D)$, a value above one favors $H_1$ over $H_0$; a value below one favors $H_0$ over $H_1$. Reversing the comparison gives $B_{01}={\mathrm{BF}}_{01}=1/{\mathrm{BF}}_{10}$, the opposite direction used in some formulas below.

For mutually exclusive, exhaustive $H_0,H_1$ with positive prior probabilities, divide their two Bayes' theorem formulas. Their common data normalizer cancels, leavingThe Bayes factor therefore multiplies prior odds into posterior odds. It is evidence conditional on the specified statistical models and their within-model prior distributions; it is not $\mathbb{P}(H_1\mid D)$. Model prior probabilities supply a separate input.^[ref]

For an original stipulated example, consider ten independent and identically distributed Bernoulli trials, with heads as success. Compare $H_0:p=0.4$ with $H_1:p=0.7$, both fixed before seeing the data. Take $D$ to be one particular ordered sequence containing eight heads and two tails. The likelihood function under either point model is $p^8(1-p{)}^2$, soIf $\mathbb{P}(H_1)=0.1$, the prior odds for $H_1$ against $H_0$ are $1/9$. Hence the posterior odds are $21.9909706/9\approx 2.4434412$, andEqual model prior probabilities would instead give ${\mathrm{BF}}_{10}/(1+{\mathrm{BF}}_{10})\approx 0.956505$. The same Bayes factor can thus accompany different posterior probabilities. These statements condition on the two-model comparison; they do not establish that either statistical model describes a real coin.

Using the heads count $K=8$ instead of the ordered sequence changes the sample space. The binomial distribution assigns it mass $\left(\genfrac{}{}{0ex}{}{10}{8}\right)p^8(1-p{)}^2$. The same binomial coefficient multiplies both models' Bayesian model evidence and cancels in their Bayes factor. Mixing an ordered-sequence numerator with a count denominator would introduce an erroneous factor of $45$.

Fix the candidate statistical models, their parameter prior distributions, and model prior probabilities before using the current data for this update. The Bayes factor is useful because it separates evidence from model prior odds, but it still depends on the parameter prior distributions. An improper prior has an arbitrary multiplicative normalization: multiplying its kernel by $c$ multiplies its formal marginal likelihood by $c$, even if the within-model posterior distribution is normalizable. Without a separately justified convention, the ordinary Bayes factor is then undefined up to arbitrary constants.^[ref]

For a finite measurable partition $(B_j)$ with positive probabilities, and an event $A$ of positive probability,Its beauty is that a model predicting observations from a hidden state can also describe uncertainty about the state after an observation. Each numerator is a joint probability; the denominator, supplied by the law of total probability, makes the answers sum to one. This is a direct consequence of conditional probability, rather than an extra axiom.

Select the coin from law of total probability once, then toss that same coin repeatedly, with conditionally independent tosses given its identity. Let $B$ denote the $3/4$-head coin, initially of prior probability $1/5$. After one head $H$, its posterior probability isThe likelihood of a head is larger under $B$, but $B$ remains less likely than the other coin because its prior probability was smaller. A large likelihood for an observation does not itself imply a large posterior probability for the hypothesis.

After an ordered sequence containing $h$ heads and $t$ tails, its likelihood ratio between the two coins is $(3/2{)}^h(1/2{)}^t$. Bayes' theorem therefore gives posterior oddsEach head multiplies the odds by $3/2$ and each tail by $1/2$. Three heads give posterior probability $27/59$, still below one half; four give $81/145$, above one half. The finite-model update and the distinction between prior probability, likelihood and posterior probability are set out in MIT 18.05, Reading 11: Bayesian Updating with Discrete Priors.

The hidden selection also creates dependence. Each toss has marginal head probability $11/20$, but two heads have joint probability $(4/5)(1/2{)}^2+(1/5)(3/4{)}^2=5/16$, greater than $(11/20{)}^2=121/400$. Their indicator covariance is $1/100$. Seeing a head makes the head-favoring coin more likely, so it also raises the chance of a future head. This is covariance induced by a shared latent variable, despite conditional independence of the tosses given the coin.

Historically, Thomas Bayes's essay on an unknown success probability appeared in 1763 after his death, communicated by Richard Price. Its opening problem asks how observed successes and failures constrain that unknown probability; Price's introductory letter explains the connection to reasoning from experience. That is a more substantial inference problem than simply rearranging today's elementary identity. ^[ref]

Chernoff information measures how quickly repeated independent evidence can distinguish two fixed probability distributions. Suppose a statistical hypothesis test observes $n$ independent and identically distributed random variables, with common law $P$ under one simple hypothesis and $Q$ under the other. Give the two hypotheses fixed positive prior probabilities, and let $p_n^{\ast }$ be the smallest achievable average error probability. For distinct positive probability mass functions on a common finite alphabet,Here $\ln$ is the natural logarithm, so the error exponent is measured in nats per observation. This optimal rate is unchanged by choosing different fixed positive prior probabilities; the best decision rule and finite-sample error can still change.^[ref]

The interpretation is $p_n^{\ast }=\exp (-nC(P,Q)+o(n))$: a larger Chernoff information means fewer observations are needed at the leading exponential scale. It does not assert the exact finite-sample identity $p_n^{\ast }=e^{-nC(P,Q)}$. The finite-sample Chernoff error bound and the explicit Chernoff information for complementary Bernoulli distributions make this distinction concrete. This is useful when designing repeated measurements or comparing sensors whose output probability distributions differ under two candidate conditions.

For probability mass functions with zero entries, define the affinity for $0<s<1$ and use $C(P,Q)=-\ln {\inf }_{0<s<1}{\sum }_{x}P(x{)}^{s}Q(x{)}^{1-s}$, taking endpoint limits when needed. This avoids an ambiguous $0^0$ convention. Identical laws give $C=0$; disjoint probability supports give $C=\infty$, because one observation separates them perfectly. The displayed finite-rate theorem concerns fixed laws with common positive support, not laws that change with $n$ or arbitrarily dependent repeated observations.

Let $P$ be the Bernoulli distribution with success probability $0.2$ and $Q$ the Bernoulli distribution with success probability $0.8$. For $0\le s\le 1$,The two summands have product $0.16$. Their arithmetic mean is at least their geometric mean, with equality at $s=1/2$. ThusThe same argument gives $C=-\ln (2\sqrt{p(1-p)})$ for complementary parameters $0<p<1$. At $p=1/2$ the laws agree and $C=0$.

With equal prior probabilities, count the successes $K$ among $n$ observations. The likelihood ratio is $Q^{\otimes n}/P^{\otimes n}=4^{2K-n}$, so choose $Q$ when $K>n/2$ and $P$ when $K<n/2$. If $n$ is even, either decision or a fair tie randomization gives the same average error at $K=n/2$. Under $P$, $K$ has the binomial distribution with parameters $n,0.2$. Symmetry therefore gives the exact optimal average errorThe last term applies only for even $n$; it represents half the tie probability. Direct finite sums give $p_1^{\ast }=0.2$, $p_5^{\ast }=0.05792$ and $p_{20}^{\ast }\simeq 0.00157912$. Twenty observations reduce the optimal average error to about 0.158%. The finite-sample Chernoff error bound at $n=20$ is $\frac{1}{2}(0.8{)}^{20}\simeq 0.00576461$, a guarantee rather than the exact error.

Reducing the exact error from $p_{20}^{\ast }$ to $p_{50}^{\ast }\simeq 1.29363\times 10^{-6}$ illustrates the value of repeated evidence. The example assumes a single fixed coin law throughout the experiment. If repeated sensor readings share an unmodelled common error, increasing their number need not supply this much independent information.

A normative Bayesian inference model of retrospective confidence. For two exhaustive alternatives with equal prior probabilities, let $L=\log [p(E\mid H_1)/p(E\mid H_0)]$ be the log-likelihood ratio. Bayes' theorem gives posterior odds $e^L$. Choosing the alternative with the larger posterior yields $P(\text{correct}\mid E)=(1+e^{-|L|}{)}^{-1}$. Unequal priors add prior log odds; choosing a different action changes which posterior probability is appropriate. This calculation does not establish that an observer knows the true likelihoods.

Fix $r>1$ ravens and $M>1$ nonblack objects in a finite population. Under $H_0$, all ravens are black; under $H_1$, exactly one is nonblack. The total numbers $r,M$ are held fixed under both hypotheses, so the classification of one nonraven also changes between hypotheses. This is a deliberately specified model for Bayesian confirmation, not a universal rule about how bird observations supply evidence.

A uniform draw among nonblack objects returns a nonraven with probabilities $1$ under $H_0$ and $1-1/M$ under $H_1$. A uniform draw among ravens returns a black object with probabilities $1$ and $1-1/r$. The respective likelihood ratios for $H_0$ over $H_1$ are therefore $(1-1/M{)}^{-1}$ and $(1-1/r{)}^{-1}$. When $M>r$, the former is closer to one, so supplies less support under these assumptions.

With equal prior probabilities, Bayes' theorem gives posterior probabilities $1/(2-1/M)$ and $1/(2-1/r)$. More generally, with prior probability $p$ for $H_0$, replace the two expressions by $p/[p+(1-p)(1-1/M)]$ and $p/[p+(1-p)(1-1/r)]$. Previously known information or a different sampling protocol requires different likelihoods; selecting an object already known to be a nonraven need not supply new evidence.

For equal prior probabilities and independent and identically distributed observations on a finite alphabet, minimum sum of errors in a simple hypothesis test givesFor nonnegative $a,b$ and $0<s<1$, $\min (a,b)\le a^s{b}^{1-s}$. Applying this termwise and using the independence of random variables to factor the sum yieldsOptimizing $s$, with an infimum if needed, proves $p_n^{\ast }\le \frac{1}{2}{e}^{-nC(P,Q)}$. This short argument gives a bound for every $n$. The theorem identifying the optimal error exponent also needs the converse: no decision rule can achieve a strictly larger limiting exponent. A log-likelihood ratio is a sum of independent contributions, so large deviation principle supplies the relevant converse idea.

People often underweight base rates when a vivid case description competes with them, but the extent of neglect depends on representation, relevance and task. The normative benchmark is conditional on a known sampling process. If a case is randomly drawn from a population, a prior class proportion $P(H)$ must combine with the case evidence's likelihood ratio. In Bayes' theorem,Changing prior odds while holding diagnostic evidence fixed should ordinarily change posterior odds. A judgment driven only by resemblance to a class violates that benchmark when the stated premises apply.

In Daniel Kahneman and Amos Tversky's prediction experiments, descriptions of people were paired with different proportions of engineers and lawyers in the source group. Judgments based on an occupationally suggestive description were surprisingly insensitive to those proportions despite their relevance to a random draw. Without the individualized description, participants could use the stated proportions. This contrast is especially informative: the issue was not simply inability to read a base rate, but the way case-specific evidence displaced it. Similarity and probability can agree in ordinary situations, which helps explain why the substitution is tempting.^[ref]

The first account is the representativeness heuristic, potentially supplemented by dual-process theory. An occupational description quickly evokes similarity to a stereotypical engineer or lawyer. That judgment does not inherently encode class prevalence or evidential reliability, so it can dominate the probability estimate. More controlled reasoning can integrate the base rate, but may fail to engage, misunderstand the statistical structure or fail to override the compelling response. This account predicts errors especially when representativeness and prior probability conflict, with possible improvement when people deliberately compare alternative hypotheses and their frequencies.

It is too strong to infer from a biased answer that no statistical information was processed. Wim De Neys and colleagues contrasted problems where descriptions and base rates conflicted with control problems where they agreed. Although verbal reports often emphasized the description, measures including decision time, later memory for the base rates and reconsideration behavior indicated sensitivity to the conflict. Statistical information can therefore influence processing without determining the final answer. This challenges a simple account in which analytic monitoring is always absent; it is compatible with incomplete integration or unsuccessful correction.^[ref]

The second account emphasizes ecological rationality and information format. Gerd Gigerenzer and Ulrich Hoffrage argued that algorithms adapted to natural frequencies can perform Bayesian updating more easily than algorithms given separately normalized probabilities. Natural sampling represents the relevant nested sets directly: how many cases belong to each class, and how many within each class produce the evidence. The denominator can then be the whole set of evidence-positive cases, without reconstructing joint counts by multiplying conditional probabilities.

Their 1995 experiments compared probability and frequency versions of Bayesian problems while also varying the information menu. Frequency formats substantially increased the use of Bayesian solutions across the problem set, assessed using solution processes as well as final answers. This supports a representational contribution to apparently poor reasoning; it does not show that everyone is spontaneously Bayesian or that every frequency phrasing helps. The preserved nesting and common reference population matter. Normalizing both classes to independent “out of one hundred” samples can remove precisely the base-rate information the natural-frequency representation was meant to preserve.^[ref]

An original numerical illustration makes the distinction explicit. Suppose a randomly sampled population contains $200$ category-$H$ cases and $800$ non-$H$ cases. A cue occurs in $75\%$ of the first class and $25\%$ of the second. There are then $150$ cue-positive $H$ cases and $200$ cue-positive non-$H$ cases, so $P(H\mid E)=150/(150+200)=3/7$, approximately $0.429$. Reporting $0.75$ would confuse the cue's conditional rate in class $H$ with the probability of class $H$ given the cue. These are constructed teaching counts, not results from a cited experiment.

The two accounts agree that the task representation matters and that naïve answers need not obey the normative model. They disagree in emphasis: the heuristic account highlights substitution and control, while the ecological account highlights a mismatch between the supplied representation and an available inference algorithm. Frequency success alone cannot decide between them, because a clearer nested-set display can also make deliberate correction easier. Conversely, conflict sensitivity constrains an absolute “base rates were never noticed” story without proving successful Bayesian computation.

A discriminating design would cross numerical format with description–base-rate conflict, preserve the sampling assumptions and information content, and measure answers together with process indicators. It should distinguish a failure to encode prevalence, a failure to combine encoded evidence and a rational belief that the stated reference group is irrelevant. Verbal descriptions seldom specify a likelihood ratio, so prior-insensitivity is easier to diagnose than a unique numerical Bayesian answer for every description. If descriptions were selectively rather than randomly sampled, or a base rate were not applicable to the individual, the simple benchmark would change. Accordingly, base-rate neglect is a reproducible conditional pattern, not evidence that people invariably ignore statistics or possess no Bayesian competence; both heuristic competition and representational structure help explain when it occurs.

The Metropolis–Hastings algorithm proposes $y\sim q(\cdot \mid x)$ and moves from the current state $x$ with acceptance probabilityOn rejection it retains $x$. The target need only be known up to its normalizing constant, but it must define a proper probability distribution. The accepted off-diagonal probability flow iswhich is symmetric in $x,y$ and proves detailed balance. This permits considerable flexibility in proposals, including joint moves when a posterior distribution has strongly correlated parameters, but requires proposal tuning and can waste computation on rejections.

A Gibbs sampler instead updates a coordinate or block from its exact full conditional distribution, holding the other coordinates fixed. This is a special Metropolis–Hastings algorithm update with acceptance probability one: the joint density factors into the unchanged marginal density and the proposed conditional density, which cancel in the acceptance ratio. The coordinate update preserves the target; composing such updates therefore preserves it too, although an entire deterministic sweep need not itself be reversible.

Use Gibbs sampling when the conditional laws are easy to simulate and suitably chosen blocks mix well. Scalar updates can mix poorly under strong posterior distribution correlation, whereas a blocked Gibbs sampler can alleviate that dependence. Prefer a flexible Metropolis proposal when conditional draws are difficult or a well-tuned joint proposal is more effective. Variations include random or systematic coordinate scans, random-walk Metropolis algorithms, independence proposals, and using Metropolis updates for difficult blocks inside a Gibbs scheme. Both methods generate dependent draws; usable long-run estimates require an appropriate irreducible, aperiodic chain and assessment of mixing, not merely a formal update rule.

For the change-point calculation, put $v={\sigma }^2$ andMultiplication of the Gaussian likelihood function by the specified prior distribution gives the formal posterior distribution kernel, with respect to $d\theta d\varphi dv$ and counting measure on $m$,

There is an empty component under an improper prior at $m=n$. The sum $S_n$ is independent of $\varphi$, whose prior distribution is flat on all of $\mathbb{R}$. For every fixed finite $\theta$ and $v>0$, the kernel is a strictly positive constant in $\varphi$, henceThe configuration has positive prior probability $1/n$. Integrating over, for example, a bounded interval of $\theta$ and $1\le v\le 2$ already makes the joint normalization infinite. Thus the printed prior distribution does not produce a posterior probability distribution. This failure is present for every data set, not just for a special arrangement of observations.

The formal full conditionals expose exactly where the requested sampler breaks. For $m<n$, completing the two mean squares givesThe variance kernel is that of $\mathrm{IG}(n/2,S_m/2)$ when $S_m>0$, and the discrete conditional probabilities are proportional to $e^{-S_m/(2v)}$. At $m=n$ the conditional kernel for $\varphi$ is constant and cannot be normalized; it is not a normal distribution with an infinite variance. Therefore no valid Gibbs sampler, and no value of $P(m=n\mid x)$, exists for the model as printed. A chain that simply substitutes an arbitrary distribution for this missing conditional would be sampling a different model.

Here is a fully specified proper-prior repair retaining the no-change configuration. Choose fixed hyperparametersindependently, with $t_{\theta },t_{\varphi },a_0,b_0>0$. Its Gaussian change-point posterior with proper priors isIt is normalizable even for degenerate data: the likelihood function is at most $(2\pi {)}^{-n/2}{v}^{-n/2}$, and its integral against the variance prior distribution is bounded byThe normal mean priors integrate to one, and the normalizing constant is positive. This proves posterior propriety for the repaired model.

Completing the conditional squares gives the Gaussian change-point Gibbs updates. DefineThe four updates areAt $m=n$, $M_{\varphi }={\mu }_{\varphi },V_{\varphi }=t_{\varphi }$, giving its proper prior distribution unchanged. Starting from any finite means, positive variance and allowed $m$, cycle through these four updates using the latest values. The sum-of-squares terms can be evaluated from cumulative data sums; normalize the discrete log weights after subtracting their maximum to avoid numerical underflow. The proper positive conditionals allow movement among all split values and throughout the continuous parameter support.

Under this specified repair, $P(m=n\mid x)$ is the posterior probability of no change within the observed sequence: all $n$ data points belong to the first mean regime. For $N$ retained draws after initialization has ceased to affect the estimates, its Markov chain Monte Carlo estimator isAlternatively average the displayed conditional probability of $m=n$ at the sampled continuous parameters, giving an estimate by Rao-Blackwellization. Autocorrelation must be accounted for in its Monte Carlo uncertainty. These are probabilities and estimators for the explicitly repaired model; the original unnormalizable kernel supplies neither.

An alternative that allows small departures from chance, while assigning little probability to near-perfect prediction or near-perfect failure, is more plausible than a uniform prior distribution. For example, a symmetric beta distribution $\mathrm{Beta}(a,a)$ with $a>1$ is centered at $1/2$, withChoosing $a=50$ gives a prior standard deviation of about $0.05$ and concentrates on modest deviations. Its concentration should represent substantive beliefs specified before examining this experiment, rather than being tuned to the observed count. A continuous prior distribution centered at $1/2$ still places no point mass at $H_0$; the separate model prior probability supplies that mass.

The reversible-jump Markov chain Monte Carlo state comprises the model index and that model's parameter. Its unnormalized posterior density in model $j$ isFor models of equal dimension, choose model $k\ne j$ with probability $s_{jk}(\theta )$ and propose ${\theta }^{\prime }$ using density $q_{jk}({\theta }^{\prime }\mid \theta )$ on the destination parameter space. The equal-dimension reversible-jump acceptance probability isOn acceptance change both model and parameter; on rejection retain both. Choose model proposals connecting all models of positive posterior mass and within-model kernels exploring their supports, with aperiodicity, to obtain an ergodic chain. Include within-model updates targeting its conditional posterior, and use model occupation proportions after the initial transient to estimate posterior model probabilities. The formula includes the model prior probabilities, parameter prior densities, likelihoods, reverse model-selection probability, and reverse parameter-proposal density.

This direct-density formula is Metropolis–Hastings algorithm on a disjoint union of equal-dimensional spaces. If instead a move uses a deterministic bijection $T:(\theta ,u)\mapsto ({\theta }^{\prime },u^{\prime })$ with auxiliary proposal densities $g_{jk}(u\mid \theta )$ and $g_{kj}(u^{\prime }\mid {\theta }^{\prime })$, useThe Jacobian determinant is one for identity matching, but equal model dimensions alone do not make a nonlinear map volume-preserving. For a direct proposal density the change of variables is already included in that density, so no additional Jacobian factor is inserted.

For the two Poisson models, use the shape-rate convention for the gamma distribution:If $b$ denotes scale instead, replace every occurrence of the rate $b$ below by $1/b$. Introduce the inactive ${\gamma }_2$ under model $2$ with the proper pseudo-prior $\pi ({\gamma }_2)$, independent of ${\gamma }_1$. This pseudo-prior augmentation for model comparison does not change model $2$'s marginal likelihood because the inactive density integrates to one. The two augmented targets on the same positive quadrant arewhereUse a symmetric cross-model proposal that flips the model label and leaves both parameters unchanged. The matching map is the identity, with unit Jacobian determinant, and the common prior factors cancel. Hence the model-switch acceptance probabilities areOnly the second observation's factor changes because ${\gamma }_1$ is retained as its shared candidate mean. Evaluate the ratio on a logarithmic scale as ${\gamma }_2-{\gamma }_1+x_2(\log {\gamma }_1-\log {\gamma }_2)$ for the first direction.

For within-model moves, Poisson-gamma conjugacy gives exact Gibbs sampling refreshes:The two draws are conditionally independent within either model. Choose with fixed positive probabilities between this block refresh and the cross-model proposal. Both moves preserve the augmented posterior; refreshing the inactive parameter maintains its correct distribution and supports mixing. The fraction of retained states labelled $M_j$ estimates $\mathbb{P}(M_j\mid x)$.

An equivalent literal dimension-changing implementation deletes ${\gamma }_2$ on a proposed $1\to 2$ move, and on a $2\to 1$ move draws an auxiliary $u\sim \mathrm{Gamma}(a,b)$ and sets ${\gamma }_2=u$. The birth matching has $1+1=2+0$ dimensions and unit Jacobian determinant. Its proposal density cancels the new parameter's prior density, giving the same acceptance ratios above for symmetric selection of move directions. This provides suitable moves without storing an inactive parameter.

For an independent check of Poisson mean equality model comparison, the exact Bayes factor isIt follows by integrating the two gamma likelihood kernels; the common factorial terms cancel. With equal model priors, the exact posterior probability of $M_2$ is $m_2/(m_1+m_2)$, which can also be used to check the RJ-MCMC occupation estimate.

With equal model prior probabilities, Bayes factor updating gives $w_0=B_{01}/(1+B_{01})$ and $w_1=1/(1+B_{01})$. The Bayesian model averaging posterior density is thereforeThis is a two-component mixture model of the normal distributions already derived. For the numerical observation above, $w_0\simeq 0.10829$ and $w_1\simeq 0.89171$. Consequently it is predominantly the wide-model posterior, not predominantly the narrow one.

With equal priors, the probability of a correct result from this quantum measurement isThe two required amplitudes areThereforeIts maximum occurs at $\beta =-\pi /8$ and is

Here $|⟨a_0|a_1⟩|=1/\sqrt{2}$, so $\theta =\pi /4$ and $\sin \theta =1/\sqrt{2}$. The stated Helstrom-Holevo bound is consequently exactly the same value, and the basis above is the Helstrom measurement for the zero and plus states.

In quantum channel discrimination, prepare a density operator ${\rho }_{HR}$ on the channel input $H$ and an optional quantum ancilla $R$. Under hypothesis $j\in \{1,2\}$ the output isUse a two-outcome quantum measurement $\{Q,I-Q\}$ and decide for $T_1$ on outcome $Q$. The conditional Type I error and Type II error areWith prior probabilities $p$ and $1-p$, symmetric Bayesian discrimination minimizes the average error $p\alpha +(1-p)\beta$ over the input and measurement.

By symmetry, each student in the university has probability $(S/N)(1/S)=1/N$ of being in the college and then selected. Hence the prior probability is

The prior probability is $1/10$. Using Bayes' theorem, the posterior probability after two absences isThus the probability rises from 10% to about 73.5% within the assumed attendance model. Two absences are 25 times as likely under the lazy model, overcoming the prior odds of $1:9$; they still do not make the classification certain.

The raven paradox combines a plausible equivalence principle with an overly unrestricted account of positive-instance scientific confirmation. Let $R(x)$ and $B(x)$ mean that $x$ is a raven and black respectively. ThenA white shoe is a positive instance of the right-hand generalization. If every positive instance confirms its universal generalization, and logically equivalent statements receive the same evidential support, inspecting the shoe seems to confirm the original raven claim. There is no contradiction in the equivalence itself. The tension concerns the evidential rule and the intuition that an arbitrary shoe teaches us nothing about birds.^[ref]

The right response is to make background knowledge and sampling explicit. Under Bayesian confirmation, evidence increases support for $H$ precisely when its likelihood is greater under $H$ than under the alternative, assuming nondegenerate prior probabilities. A fact can therefore be compatible with $H$ yet provide no new support. The mere existence of a white shoe does not determine the needed likelihood ratio.^[ref]

Consider a deliberately simple finite model. There are $N$ objects, exactly $r$ ravens, and $M=N-r$ nonravens, with $M>r>1$. All nonravens are white. Under $H_0$ all ravens are black; under $H_1$ exactly one raven is white. Draw uniformly from the entire population. The event of drawing a white nonraven has probability $M/N$ under either hypothesis. Its Bayes factor is one and it leaves the prior odds unchanged. A black raven instead has likelihoods $r/N$ and $(r-1)/N$, so gives Bayes factor $r/(r-1)$ for $H_0$.

Now change the protocol: draw uniformly from the white objects, then discover that the selected object is a nonraven. This result has likelihoods $1$ and $M/(M+1)$. It gives the small positive Bayes factorFor $N=1000$ and $r=10$, this is $991/990$, whereas sampling a black raven from the full population gives $10/9$. The white nonraven can now help because the search could have revealed a white raven. The two experiments have different selection information, so their differing evidential values do not violate logical equivalence.

This model establishes possibilities, not a universal numerical law about shoes. Different alternatives, population information or selection procedures change the result. If the shoe's being a nonraven was already known, discovering that it is white need not add anything. Conversely, ruling out a potential counterexample can supply a small amount of information. Retain logical equivalence, but reject the unconditional instance-confirmation rule. Once the complete evidential situation is specified, neither zero support nor weak positive support is paradoxical, and neither amounts to proof of a universal law.

General philosophical scepticism has important implications for the standards and limits of scientific justification. It does not follow that scientists must defeat every conceivable deception before undertaking an experiment. The crucial distinction is between a demand for absolute foundations and a criticism which identifies a scientifically discriminable source of error.

Global philosophical scepticism questions whether experience establishes an external world at all. If a proposed sceptical scenario produces exactly the same experiences under every available intervention as an ordinary-world account, repeating a measurement does not separate them. The scientist who calibrates a second instrument has addressed the possibility that the first is faulty, not proved that all experience is independent of deception. Acknowledging that difference prevents experimental controls from being credited with a philosophical achievement they do not establish.

The problem of induction bears more directly on prediction. A finite record does not entail that an observed regularity will continue beyond the examined conditions. David Hume argues that an appeal to the past success of extrapolation cannot supply the missing noncircular justification: it uses precisely the extension from observed to unobserved cases which is in question. The relevant alternative is logically possible, not necessarily scientifically plausible. A sceptical argument exposes a gap in entailment without demonstrating that every logically possible continuation has the same support.^[ref]

For a physical example, measurements may establish a nearly linear relation between electrical resistance and temperature over a limited interval. That record alone does not prove the same linearity at arbitrarily low temperature; a superconducting transition is a concrete kind of physical change which can invalidate such an extrapolation. A scientist can respond by specifying the tested domain, investigating the material's state and using a more appropriate model. Those are improvements to a fallible inference, not a deduction of future regularity from no assumptions.

Similar care applies to instruments. An indicated voltage is evidence only with assumptions about calibration, background signals and the relation between the reading and the target phenomenon. Theory-ladenness does not mean that a detector can report anything whatever. Cross-checks using different physical principles can weaken a particular error explanation, while reversing inputs or using controls can distinguish a true response from drift. Such procedures make doubts specific and testable. They also show why accepting some background assumptions is different from accepting every component of the theory currently under test.

One response to the general challenge is fallibilism. Scientific warrant need not amount to an impossibility of error. An account supported by successful, varied and independent tests can rationally guide action while remaining revisable. Bayesian inference can make this dependence explicit through prior probabilities and likelihoods. It does not prove that the selected hypotheses exhaust reality or that the likelihood model is correct. Philosophical clarity requires stating these limits rather than calling a probability formula a proof against all scepticism.

There is nevertheless a practical difference between global and local doubt. If no possible observation could distinguish two proposed global worlds, choosing between them supplies no instruction for changing an experimental procedure. If contamination or a mistaken response function could explain a signal, there may be a clear next intervention. Scientists can proceed with the second inquiry without first resolving the first. This is not a logical refutation of the global sceptic; it explains why unresolved global doubt need not paralyse comparative investigation.

Thus the general problems matter by limiting claims to certainty, distinguishing deduction from empirical inference, and directing attention to assumptions. Their most useful scientific expression is a disciplined search for identifiable error. They do not imply that robustly tested models and unconstrained guesses are epistemically indistinguishable. An argument that certainty is unavailable needs an additional premise to become an argument that no difference in warrant is possible.

Bayes' theorem solves a precise part of the scientific confirmation problem: it gives a coherent relation between prior commitments, evidential likelihoods and updated probabilities. It does not select those commitments from no premises, and it should not be identified with every meaning of confirmation.

For background information $K$, a hypothesis $H$ and newly considered evidence $E$, let $p=P(H\mid K)$, $a=P(E\mid H,K)$ and $b=P(E\mid \neg H,K)$. With a positive denominator,The first formula requires a positive denominator; the odds formula additionally assumes $0<p<1$ and $b>0$. The second equation separates prior odds from the Bayes factor. For $0<p<1$, the evidence raises the probability of $H$ exactly when $a>b$. It is not enough that $E$ is likely under $H$; it must be more likely than under the relevant alternative mixture. This guards against an inverse-probability error in evaluating scientific tests.

For example, let the prior be 0.01 and the two likelihoods be 0.8 and 0.08. The Bayes factor is ten, but the posterior probability is only $10/109\simeq 0.0917$. The evidence strongly favours $H$ relative to its previous odds without making $H$ probable in an absolute sense. Evidence can therefore confirm a hypothesis while leaving a rational investigator doubtful. The comparison also depends on which alternatives are included in $\neg H$ and how their probabilities are distributed.

The raven paradox shows another benefit. Logically equivalent claims receive identical probabilities in the same probabilistic model, but this does not force every investigation of a white object to provide the same evidence about ravens. If an object is selected because it is already known to be a shoe, observing that it is not a raven may have equal likelihood under the rival hypotheses and provide a Bayes factor of one. Random selection from a specified population can give different likelihoods. The sampling procedure and background information must be represented, rather than letting a grammatical positive-instance rule determine evidential support.^[ref][ref]

The problem of old evidence is a genuine limit of the simplest updating account. If $E$ already belongs to $K$, conditioning on it again produces no increase. Yet an old measurement can be important support for a newly formulated theory. One may assess support relative to a specified earlier or counterfactual background, or model learning a previously unrecognized connection between theory and evidence. These are substantive additions. Simply applying the same conditioning rule to an already certain fact does not represent the new intellectual achievement.^[ref]

Other limitations concern model construction and priors. A hypothesis assigned zero prior probability remains at zero after ordinary conditioning on evidence of positive total probability. A badly specified likelihood can generate confident but unreliable results. Different reasonable priors need not disappear under finite or non-discriminating evidence. Nor does the theorem supply a noncircular solution to the problem of induction: the model's connection to future physical events remains an empirical commitment.

Thus the Bayesian framework gives a rigorous account of probability change, comparison of alternatives and some apparent paradoxes. These are real achievements. It is incomplete as a general recipe for scientific judgment unless the hypotheses, backgrounds, likelihoods and forms of learning are also defended. A mathematical rule of coherent updating can organize evidential reasoning without establishing that every coherent input represents nature correctly.

Numerous successful tests can justify a fallible belief in a general hypothesis, but do not usually entail its truth. If finitely many tested instances satisfy $F(x)\Rightarrow G(x)$, a further untested instance can still fail. The logical gap is the problem of induction, not a practical complaint that the sample was merely too small. Appealing to the past success of inductive reasoning to establish its universal reliability uses the disputed form of inference itself.^[ref]

The kind of testing matters as much as the count. Repeating one measurement with the same unnoticed systematic bias can add little. Independent apparatus, novel circumstances and predictions on which competing theories disagree can be much more informative. A test deliberately designed to expose a failure is stronger than a loose resemblance noticed after the outcome. Still, auxiliary hypotheses and instrument assumptions must be considered when interpreting either success or failure.

Bayesian confirmation expresses the distinction quantitatively. For independent trials with success probability $p$, $n$ successes have likelihood $p^n$. In a simple comparison between a certain-success hypothesis and a specified alternative $p=q<1$, the Bayes factor is $q^{-n}$. This favours the first hypothesis without proving it. Independence, the alternative and the prior probabilities are substantive assumptions; repeated correlated results cannot automatically be multiplied as independent evidence.

There is also a difference between a universal statement and an adequate approximation within a domain. Many successful low-speed tests of Newtonian mechanics justify its use there, without establishing its unrestricted truth at relativistic speeds. A reasonable conclusion is fallibilism: well-designed and varied successful tests can warrant acceptance or high confidence, while the scope and remaining uncertainty should stay explicit. If “justified” means deductively certain, finite testing generally does not suffice; if it means rationally supported, it often can.