The Gold Standard of Uncertainty Quantification | When a Valid Correlation Matrix Cannot Fit Binary Inputs

Let B₁, B₂ and B₃ take values zero and one, with each equal to one with probability one half. Think of three recorded threshold events in one trial. The labels describe whether each event occurred, without claiming that the events are independent or that a physical mechanism has been identified.

A complete trial contains all three values together. Observing B₁ in one population and B₂ in another does not automatically produce a joint observation. The intended population, time window and event definitions must be shared before the dependence quantities can be interpreted as one model.

Each input has mean one half and variance one quarter. These marginal facts are part of the question, not adjustable conveniences. If a simulator produces the desired correlations by generating continuous values, it has not solved the stated binary problem. We need a probability allocation over the eight permitted triples of zeros and ones.

Contents · Next chapter

Write a = Corr(B₁,B₂), b = Corr(B₁,B₃) and c = Corr(B₂,B₃). The correlation matrix has rows [1,a,b], [a,1,c] and [b,c,1]. Every genuine finite-variance joint distribution produces a positive semidefinite matrix: each weighted sum has nonnegative variance.

The existing guide on testing whether pairwise correlations form one valid joint model derives that requirement and the three-input determinant condition. Retain that check. The present lesson adds the support and marginal constraints that a matrix alone does not express.

Positive semidefiniteness is necessary for our binary model. It is sufficient for constructing some appropriately standardised real-valued variables with the supplied matrix, but that broader existence statement does not insist that those variables take only two values with the stipulated probabilities. The words some variables and these binary variables refer to different feasibility questions. Keeping them separate prevents a successful matrix test from being asked to prove more than it actually does.

Contents · Previous chapter · Next chapter

Set all three proposed correlations equal to −0.45. An equal-correlation three-by-three matrix has one eigenvalue 1 + 2r and two eigenvalues 1 − r. Here they are 0.10, 1.45 and 1.45, all positive. Thus the proposal is positive definite, rather than merely sitting on a singular boundary.

One can check the special eigenvalue directly: multiplying the matrix by (1,1,1) gives 0.10 times that vector. Any vector whose entries sum to zero receives multiplication by 1.45. These directions span the space and establish the complete eigenvalue account.

Each separate binary pair is also feasible. For fair binary inputs, a correlation of −0.45 corresponds to a both-one probability (1 − 0.45)/4 = 0.1375. The both-zero probability is also 0.1375, with the two discordant probabilities each 0.3625. All four are nonnegative and sum to one. The contradiction appears only when the three pairs must coexist.

Contents · Previous chapter · Next chapter

Let N = B₁ + B₂ + B₃, the number of events occurring in one trial. Its mean is 1.5. Each variance contributes 0.25, and each pairwise covariance is r/4. Therefore Var(N) = 0.75 + 1.5r.

At r = −0.45, this formula gives variance 0.075. That number is positive, so a nonnegative-variance check alone sees no contradiction. But N can only equal zero, one, two or three. Its distance from its mean 1.5 is always at least 0.5. Consequently (N − 1.5)² is at least 0.25 on every possible trial.

Taking expectations gives Var(N) ≥ 0.25, contradicting 0.075. The proposed binary model is impossible even though the matrix is positive definite. The missing constraint came from the integer-valued support, not from another eigenvalue. A continuous variable could vary in a narrow band around 1.5; this event count cannot occupy that band.

Contents · Previous chapter · Next chapter

The same count argument gives 0.75 + 1.5r ≥ 0.25, hence r ≥ −1/3. The matrix condition alone allows r ≥ −1/2. The interval from −1/2 up to −1/3 therefore contains matrices that are valid as correlation matrices but incompatible with these three fair binary margins.

At r = −1/3, N must equal one or two with total probability one, since any mass at zero or three would raise the squared deviation above its minimum. Equal marginal probabilities additionally give equal total mass to N = 1 and N = 2.

A construction makes the boundary concrete: assign equal probability one sixth to 001, 010, 100, 011, 101 and 110. Every input equals one in three of the six patterns. Each pair is simultaneously one in exactly one pattern, so its correlation is 4(1/6) − 1 = −1/3. This is a valid boundary model with no negative or missing probability mass.

Contents · Previous chapter · Next chapter

Define Xᵢ = 2Bᵢ − 1. Each Xᵢ takes values minus one and plus one with equal probability, has mean zero and variance one. Positive affine rescaling preserves the pair correlations, so E(X₁X₂) = a, E(X₁X₃) = b and E(X₂X₃) = c.

This representation makes individual joint patterns easier to inspect. If two signs agree, their product is plus one; if they disagree, it is minus one. In a triple there are only two possibilities: all three agree, or exactly one sign differs from the other two.

For any realised triple, the sum X₁X₂ + X₁X₃ + X₂X₃ is either three or minus one. Taking expectations immediately gives a + b + c ≥ −1. The equal-correlation bound is the special case 3r ≥ −1. This proof uses the actual pattern geometry rather than a determinant, and prepares us to derive the remaining necessary inequalities.

Contents · Previous chapter · Next chapter

Group each sign triple with its complete reversal. The all-equal class contains +++ and −−−. The second class contains ++− and −−+. The third contains +−+ and −+−. The fourth contains +−− and −++.

Within these four classes, the pair-product vectors (X₁X₂,X₁X₃,X₂X₃) are respectively (1,1,1), (1,−1,−1), (−1,1,−1) and (−1,−1,1). Reversing all signs leaves every pair product unchanged.

Let q₀ through q₃ be the total probabilities of the four classes in that order. They must be nonnegative and sum to one. Correlations are their weighted averages of these four fixed vectors. A proposed triple of correlations is therefore asking for actual probabilities on four pattern types, not arbitrary coordinates inside a cube.

This compression preserves pair products while discarding how each class splits between its two reversed patterns. That discarded information will matter for other questions. For now, it provides a compact route to the feasibility test.

Contents · Previous chapter · Next chapter

The weighted averages give a = q₀ + q₁ − q₂ − q₃, b = q₀ − q₁ + q₂ − q₃, and c = q₀ − q₁ − q₂ + q₃. Combine each equation with q₀ + q₁ + q₂ + q₃ = 1.

Solving yields q₀ = (1 + a + b + c)/4, q₁ = (1 + a − b − c)/4, q₂ = (1 − a + b − c)/4, and q₃ = (1 − a − b + c)/4.

These equations reveal four necessary inequalities: every numerator must be nonnegative. There is no need to infer a negative probability indirectly through a software warning. At a = b = c = −0.45, q₀ = −0.0875, while the other three weights are 0.3625. The negative all-equal weight is a direct certificate of incompatibility. It is not a small probability that should be rounded to zero and then silently renormalised.

Contents · Previous chapter · Next chapter

Suppose the four calculated q values are nonnegative. They sum to one automatically. Choose a class with those probabilities, then select either of its reversed sign patterns with probability one half.

Each coordinate is positive in one pattern and negative in its reversal. Thus every Xᵢ has mean zero and equal sign probabilities. The class’s pair products do not change on reversal, so their expectations reproduce a, b and c exactly.

This constructs a valid fair-binary joint distribution after converting X back to B. Necessity and sufficiency have both been established for three fair binary variables. The proof does not require claiming that every feasible joint distribution must split each class equally. It only needs one valid construction whenever the weights are nonnegative.

The scope is important. Four class weights solve this particular three-input, half-probability problem. More variables create more pattern classes and additional relationships. Unequal one probabilities require preserving different marginal conditions. We will derive an eight-cell test for that changed setting instead of transporting this convenient formula outside its domain.

Contents · Previous chapter · Next chapter

Take a = 0.20, b = −0.40 and c = 0.10. Substituting gives q₀ = 0.225, q₁ = 0.375, q₂ = 0.075 and q₃ = 0.325. Each is nonnegative, and together they sum to one.

The following table gives one explicit sign-valued joint distribution by splitting each class equally. The row represents a class containing the two named patterns; each individual pattern receives half its class weight.

Reconstruct a from the class weights: 0.225 + 0.375 − 0.075 − 0.325 = 0.20. Reconstruct b and c independently before accepting the table. The explicit pattern probabilities provide a stronger existence argument than declaring the matrix numerically acceptable without showing how the binary outcomes can occur.

Contents · Previous chapter · Next chapter

For this mixed-correlation example, E(N) remains 1.5. The variance formula is now Var(N) = 0.75 + (a + b + c)/2, giving 0.70. The sum a + b + c equals −0.10.

The all-equal class has probability 0.225, so N equals zero or three with total probability 0.225. Those outcomes have squared deviation 2.25 from the mean. The other classes have total probability 0.775 and give N equal to one or two, each with squared deviation 0.25.

The direct variance is therefore 0.225 × 2.25 + 0.775 × 0.25 = 0.70, agreeing with the covariance calculation. This is a useful independent check because it moves from second moments back to actual permitted outcomes.

Neither calculation establishes which dependence pattern a real system follows. They verify that the stated finite construction has the intended properties. A simulation can be checked against these exact results, separating an implementation error from uncertainty about whether the model is appropriate for the intended application.

Contents · Previous chapter · Next chapter

Our equal reversal split provides one joint model, not necessarily the only one. Pair correlations describe expected pair products. They do not in general reveal every probability of a three-way outcome.

For a compact example, return to equal correlations r = −0.30. The all-equal class weight is 0.025 and the other three weights are each 0.325. An equal reversal split gives P(+++) = P(−−−) = 0.0125.

A family preserving fair margins and the same pair correlations can instead assign each sign triple probability [1 + r(x₁x₂ + x₁x₃ + x₂x₃) + τx₁x₂x₃]/8. Here each x is minus or plus one. Nonnegativity permits τ from −0.10 to 0.10, because the smallest fixed numerator is 0.10 in the all-equal patterns.

Consequently P(+++) can range from zero to 0.025 while all the pair correlations remain unchanged. The new calculation concerns what remains unidentified after compatibility is established. It should not be mistaken for evidence that one of those triple probabilities has been measured.

Contents · Previous chapter · Next chapter

Suppose a routine receives the equal −0.45 matrix and checks positive definiteness. It can correctly report success. There is no negative eigenvalue for a nearest-positive-semidefinite operation to remove. That operation would address the wrong missing constraint.

If a proposed revision sets all correlations to −1/3, the fair-binary model becomes feasible. But the revision has changed the dependence assumption. It has not discovered the true correlations or justified the boundary as an empirical estimate.

An alternative revision might change event definitions, marginal probabilities or the model’s permitted values. Those changes also alter the question and require explicit evidence. Replacing binary inputs with normal variables because a convenient factorisation works does not preserve the original distributional contract.

Keep the original proposal, the failed constraint, the revised inputs and the reason for the change. Then compare outputs that matter. A mathematical repair can make a model possible while leaving it poorly supported. The aim is a transparent model that can be investigated, not a successful status message detached from its meaning.

Contents · Previous chapter · Next chapter

Suppose all three correlations are reported as −0.33 after rounding to two decimal places. Under ordinary nearest rounding, a possible underlying value lies approximately between −0.335 and −0.325, with endpoint conventions depending on the rounding rule.

The equal-correlation binary boundary −1/3 lies inside that interval. Some possible underlying triples with equal values are feasible and others are not. Treating the printed −0.33 as the exact underlying value would conceal this distinction.

With three independently rounded entries, examine the four class-weight inequalities jointly. Their sums can compound small reporting differences. A useful result may be that some values within the supplied rounding intervals admit a model, while other values do not. That is a feasibility statement about an interval set rather than a recovered exact matrix.

Do not choose extra decimal places merely to obtain the desired outcome. Obtain unrounded evidence when it matters, or report the remaining ambiguity. Numerical tolerance should represent a justified computational or reporting allowance, not permission to ignore a substantively negative pattern probability.

Contents · Previous chapter · Next chapter

Now let P(B₁ = 1) = 0.20, P(B₂ = 1) = 0.40 and P(B₃ = 1) = 0.70. The inputs are still binary, but no longer have equal half probabilities. The four fair-sign class weights are no longer a sufficient description of the marginal constraints.

For any pair, write pᵢ for its one probability and qᵢⱼ for its both-one probability. Cov(Bᵢ,Bⱼ) = qᵢⱼ − pᵢpⱼ, and variance is pᵢ(1 − pᵢ). Thus a supplied correlation must first be translated into a proposed q using those actual marginal variances.

Every pair already has bounds: max(0,pᵢ + pⱼ − 1) ≤ qᵢⱼ ≤ min(pᵢ,pⱼ). These follow from the four nonnegative probabilities of both, first only, second only and neither.

A marginal probability of zero or one needs separate treatment. The corresponding binary variable is constant and has variance zero. Its ordinary Pearson correlation with another variable is undefined because the standard-deviation denominator vanishes. A supplied correlation of zero would not accurately describe that undefined quantity, even though its covariance with every finite variable is zero. Return to probabilities and cells instead of forcing a correlation formula through a zero denominator.

For the pair with probabilities 0.20 and 0.70, q lies from zero to 0.20. Its correlation range is therefore about −0.764 to 0.327. A generic minus-one-to-one check misses the restrictions created by the specified unequal margins before the third variable is even considered.

Contents · Previous chapter · Next chapter

Let t = P(B₁ = B₂ = B₃ = 1). Once the three one probabilities and the three pair intersections are supplied, every other cell can be expressed in terms of t.

The probabilities of 110, 101 and 011 are q₁₂ − t, q₁₃ − t and q₂₃ − t. The single-one patterns 100, 010 and 001 have probabilities p₁ − q₁₂ − q₁₃ + t, p₂ − q₁₂ − q₂₃ + t and p₃ − q₁₃ − q₂₃ + t.

Finally 000 has probability 1 − p₁ − p₂ − p₃ + q₁₂ + q₁₃ + q₂₃ − t. Together with the 111 cell t, these eight expressions sum to one. Reconstruct each margin by summing its four relevant cells to check the derivation.

A valid triple requires every cell to be nonnegative. This is a complete finite feasibility test for the supplied three binary margins and pair intersections. It does not assume independence, a Gaussian shape or a particular causal explanation for the events.

Contents · Previous chapter · Next chapter

Collecting the lower constraints gives t ≥ max(0, q₁₂ + q₁₃ − p₁, q₁₂ + q₂₃ − p₂, q₁₃ + q₂₃ − p₃). The upper constraints give t ≤ min(q₁₂, q₁₃, q₂₃, 1 − p₁ − p₂ − p₃ + q₁₂ + q₁₃ + q₂₃).

For the changed margins (0.20,0.40,0.70), choose intersections (0.10,0.12,0.30). The lower candidates are 0, 0.02, 0 and −0.28, so the lower bound is 0.02. The upper candidates are 0.10,0.12,0.30 and 0.22, giving upper bound 0.10.

There is therefore a complete family of compatible joint distributions with t between 0.02 and 0.10. Any value in that interval gives nonnegative cells and the stipulated margins. A particular t needs additional evidence or a declared modelling assumption. The calculation identifies possibility, not a preferred probability within the interval.

Contents · Previous chapter · Next chapter

Choose t = 0.06 in that interval. In the order 111,110,101,011,100,010,001,000, the probabilities are 0.06,0.04,0.06,0.24,0.04,0.06,0.34 and 0.16.

Their sum is one. Summing the cells where B₁ is one gives 0.06 + 0.04 + 0.06 + 0.04 = 0.20. The corresponding sums for B₂ and B₃ give 0.40 and 0.70. Pair intersections reconstruct as 0.10,0.12 and 0.30.

The covariance terms are 0.02, −0.02 and 0.02. Marginal variances are 0.16,0.24 and 0.21. The resulting correlations are approximately 0.1021, −0.1091 and 0.0891. Their moderate appearance is not the proof; the verified joint table is.

Ask the learner to hide one cell and recover it in two ways: from the total probability and from an appropriate margin. Agreement tests whether the table is being understood as one joint account. It is more informative than copying a correlation calculation without checking what outcomes the model actually permits.

Contents · Previous chapter · Next chapter

Return to three half-probability margins, but now supply every pair intersection as 0.05. Each isolated pair is feasible because its both-one and both-zero cells are 0.05, with discordant cells each 0.45.

The eight-cell interval has lower bound zero. Its final upper candidate is 1 − 1.5 + 0.15 = −0.35. Hence the upper bound is negative and no t can satisfy both ends. The pairs do not fit one three-event distribution.

This example is easy to diagnose without matrix algebra: the three events are individually common, but the proposal allows too little overlap for all three to coexist on one probability space. The contradiction belongs to their combined allocation.

Its correlation matrix also fails the ordinary matrix test, since each correlation is −0.80. That makes it a useful checking case but not the main positive-definite counterexample. Keep the two examples separate: one shows why pair tables need a global check; the −0.45 case shows why a global covariance check can still be insufficient for the required binary margins.

Contents · Previous chapter · Next chapter

Suppose a fictional output is the count N of three equally likely events. Under equal correlation r = −0.30, its variance is 0.30 and standard deviation approximately 0.548. The fair-binary construction supports these numbers.

The impossible r = −0.45 proposal would report standard deviation √0.075, approximately 0.274. That looks like a narrower uncertainty statement, but no stated binary input model can produce it. Extra digits or additional Monte Carlo draws cannot repair the missing distribution.

The appropriate response is to investigate the input contract: event probabilities, common trial definition, dependence provenance and permitted outcomes. Clipping a final variance to a convenient positive number would not restore the lost joint consistency.

Output uncertainty is downstream of input feasibility. This does not mean every model must be exhaustively enumerated by hand. It means the method used to generate or combine inputs should preserve the actual constraints, and a small exact case should be available for independent checking when the numerical workflow becomes more elaborate.

Contents · Previous chapter · Next chapter

A mathematically valid probability table need not be representable exactly by a chosen finite number of equally weighted records. In the mixed-correlation example, individual pattern probabilities are multiples of 1/80. Eighty records can therefore reproduce them exactly using counts 9,9,15,15,3,3,13 and 13 for the eight listed sign patterns.

A ten-record table cannot reproduce a cell probability 0.1125 exactly because every cell proportion is a multiple of 0.10. That is an arithmetic restriction of the finite record size, not evidence that the population model is impossible.

For probabilities without a convenient common finite denominator, a finite simulation can approximate them rather than reproduce them exactly. The resulting sample correlations fluctuate as part of that sampling process. Comparing them with a target requires acknowledging sampling variability instead of requiring exact decimal agreement. Weighted records can represent some fractions more directly, but then their common weights are part of the empirical distribution and must enter every margin and covariance consistently.

Conversely, a complete observed binary table automatically provides a possible empirical joint distribution when all rows use one common nonnegative weighting scheme. It can still be a poor estimate of another population or future condition.

Keep these questions distinct: can any population model exist, can this exact finite record represent it, and does the available record support using it for the intended prediction? A single label such as valid data cannot answer all three. The meaning of each numerical object must remain visible.

Contents · Previous chapter · Next chapter

Try a = 0.60, b = 0.50 and c = −0.10 for three fair binary inputs. Calculate all four class weights before deciding feasibility. Then compare the result with a covariance-only calculation.

The weights are 0.50,0.30,0.25 and −0.05. Thus no fair-binary joint distribution has those correlations. Yet the correlation-matrix determinant is 1 + 2(0.60)(0.50)(−0.10) − 0.60² − 0.50² − (−0.10)² = 0.32, and its two-by-two principal minors are positive. The matrix is positive definite.

For a changed version set c = 0.10. The weights become 0.55,0.25,0.20 and zero. The half-split construction gives a valid joint distribution. The absent fourth class means X₁ cannot be opposite to both X₂ and X₃ in this particular correlation pattern.

Explain the repair through the class weight rather than merely saying that the correlations became less extreme. The decisive inequality is 1 − a − b + c ≥ 0, requiring c ≥ 0.10 for these stipulated a and b.

Contents · Previous chapter · Next chapter

With more binary inputs, one can assign probabilities to every permitted joint pattern and impose margins and pair intersections as linear constraints. Nonnegative cell weights summing to one give an explicit existence formulation. An empty feasible set identifies incompatibility under the specified inputs.

The number of patterns grows exponentially with the number of inputs, so this direct description can become computationally demanding. That is a reason to choose an appropriate method and exploit justified structure, rather than assume that every positive semidefinite matrix can be generated with the requested binary margins.

Checking every triple may provide useful necessary conditions in a larger system. It should not be promoted without proof into a sufficient global test for all dimensions. The exact sufficiency we derived belongs to three fair binary variables.

A reproducible computational record should state the variables, permitted patterns, constraints, tolerances and solver result. Where possible, retain a feasible distribution as a constructive certificate. If no distribution is found, distinguish a demonstrated infeasible set from a numerical routine that merely failed to find a solution.

Contents · Previous chapter · Next chapter

Huber and Marić’s Bernoulli Correlations and Cut Polytopes, revised 30 June 2017, studies the attainable correlation vectors of symmetric Bernoulli variables and gives a constructive convex-mixture characterisation. It supplies primary context for why marginal restrictions matter beyond ordinary matrix validity.

The original calculations here derive the three-variable case directly from sign patterns and nonnegative probabilities. The equal −0.45 proposal, mixed correlations, unequal-margin cells and transfer examples are teaching constructions. They should not be attributed to observed systems or treated as datasets from the paper.

For a first reading, understand the four class weights before pursuing the wider geometric terminology. The concrete proof already explains why a negative weight is impossible and why a nonnegative mixture supplies a model. Advanced language becomes useful when it connects this small calculation to larger existence and simulation problems, rather than replacing an explanation the learner can verify.

Contents · Previous chapter · Next chapter

Start by specifying the input values, marginal probabilities and common trial. Translate correlations into pair intersections using the correct means and variances. Check the isolated pair bounds before testing global compatibility.

For three fair binary inputs, compute the four class weights. For three arbitrary binary margins, calculate the triple-intersection interval and reconstruct a complete table at a feasible value. Verify total probability, all margins and all intersections independently.

Keep the event coding alongside the matrix. Replacing B₂ with 1 − B₂ changes its marginal probability and reverses its correlations with both other nonconstant inputs. The associated joint table must be relabelled consistently. Changing only one matrix entry is not the same transformation and can break a previously feasible model. A data dictionary is therefore part of the mathematical evidence, not merely a descriptive attachment.

Then inspect what is still unidentified and which output depends on it. Save the chosen joint construction, the evidence or assumption selecting it, and the numerical procedure used to propagate it. A fixed seed reproduces a simulation run; the distribution specification explains what that run means.

After an input change, repeat the relevant checks. A new threshold can change marginal probabilities and attainable correlations together. Keeping old correlations while changing margins can invalidate a model even when each new marginal looks sensible alone. The check follows the joint contract rather than the convenience of reusing one saved matrix.

Contents · Previous chapter · Next chapter

For repair, begin with the count N and its four permitted values. Ask why a mean of 1.5 forces squared deviations of at least 0.25. This exposes the missing support constraint without requiring familiarity with eigenvalues.

For stabilisation, calculate class weights for several proposals and construct one valid table. The learner should regenerate the correlations from their own table. If a weight is negative, ask which pattern class it purports to describe before considering any numerical revision.

For extension, change the marginal probabilities and derive the eight cells. Require an explanation of why a free triple intersection remains and how its bounds are obtained. Then ask for evidence that would select a particular member of the feasible family.

Progress means being able to distinguish possible, identified and empirically supported. Those are separate achievements. A learner who can name the missing constraint in a new example has gained more than someone who memorises that certain negative correlations are forbidden without explaining under which distributional assumptions.

Contents · Previous chapter · Next chapter

Is positive semidefiniteness the wrong check?

No. It remains necessary for every covariance model. The error is treating it as sufficient for arbitrary prescribed marginal distributions. Our positive-definite counterexample requires an additional binary-support check.

Does negative correlation make a binary model invalid?

No. The six-pattern boundary construction and the mixed-correlation table contain valid negative relationships. Feasibility depends on the whole proposal and its margins, not the sign of one coefficient.

Can more simulation draws make an impossible proposal accurate?

No. A simulator must generate some actual distribution. If its output appears to match an impossible target, inspect what it is truly preserving, including margins, event coding, correlations and numerical tolerance.

Does one feasible table prove the dependence is true?

It proves existence under the stated constraints. Empirical adequacy needs relevant observations and an intended-use argument. Another feasible table may answer a different three-way question differently.

Contents · Previous chapter · Next chapter

Check the covariance foundation

Return to Uncertainty Quantification and the joint correlation-matrix guide for uncertainty budgets and nonnegative variances. How Correlation Works develops the broader interpretation of association.

Preserve the complete probability question

How Joint Probability Works provides the parent framework. Statistical Independence separates pairwise and mutual independence. The rain-occurrence window guide develops a different application: why supplied time-window probabilities need the relevant event overlap.

Transfer the learning

The primary research link was checked on 25 October 2026. Continue through Gold Standard Education and How X Works. Choose three binary events, state their common observation frame and explain what would count as a genuine joint observation. Then test the model’s existence before interpreting the uncertainty it reports.

Contents · Previous chapter · Return to Uncertainty Quantification