By the correspondence theorem for ideals, an ideal of has the form with . If , apply unique factorization of ideals in a number field to . Every prime ideal dividing contains and hence . Primality gives , and maximality gives . Thus for some .
The inclusion forces : if , cancellation using the invertible fractional ideal would give , which is impossible. The zero ideal corresponds to , while the unit ideal corresponds to . Consequently the nonzero proper ideals are exactly
When , there are none, in agreement with being a field.
Put , , and
The proposed integral basis for coprime biquadratic radicands congruent to two and one is
Each entry is an algebraic integer, by the ring of integers of a quadratic field formula and closure under products. Their coefficient matrix in has determinant , so they are rationally linearly independent. It remains to show that they span every algebraic integer.
Write initially with rational . The field traces to and are and . Since , the corresponding integral basis in each subfield gives . The field trace to is . Membership in says that are integers of equal parity. As is already an integer, this forces too.
Now the field norm to is
By the ring of integers of a quadratic field formula, is an integer and has the same parity as . Thus is even, giving , and
Using , and , this becomes
Hence . Finally,
an expression with integer coefficients. This proves the asserted integral basis without assuming a general theorem about products of elements of an integral basis.
Again put and . The ring of integers of each of the three quadratic subfields has integral basis , , or , respectively: their square-free radicands are respectively congruent to modulo four. The field traces of the algebraic integer to those subfields are
They are algebraic integers, so comparison with each integral basis gives
This use of all three subfields is needed to control as well as the other coefficients.
The field norm to is
Its two coefficients must be integers. The second condition says is even, since is even. Reducing the numerator of the first coefficient modulo four gives
Reduction modulo two first gives . Since is even, both and are even. The same congruence now gives , or . Therefore every algebraic integer has the form
with four integer coefficients.
For the converse, are algebraic integers, and
Here and , so both coefficients are integers. The monic polynomial and integral dependence is transitive show that is an algebraic integer. The four displayed generators are rationally linearly independent, since their coefficient matrix in has determinant . Hence the integral basis for coprime biquadratic radicands congruent to three and two is
In this formula denotes the product of the chosen square roots; this convention also handles negative radicands.
Put , and . The hypotheses make a biquadratic extension. To check the degree of a field extension is four, if with rational , squaring gives . The case would make a rational square; the case would make a rational square. Neither can occur for the specified distinct coprime square-free integers. Hence .
The nonidentity field automorphism of fixes and sends to . Thus, for , the field norm and field trace are
If is an algebraic integer, so is its conjugate, and their sum and product are algebraic integers. Conversely, put and . The identity
shows that is integral over . If are algebraic integers, the ring generated by algebraic integers is integral over . Since integral dependence is transitive, is an algebraic integer. This proves the trace and norm integrality criterion for a quadratic extension, including elements already in .
For , the notation of part (i) gives , , and . Write . The remaining covariance is then and the new observation's remaining variance is . Thus
Equivalently,
The expression is valid also when and . For , a noiseless observation, interpreted as the limit , removes all remaining variance at the observed target.
The source's assertion that the target input always gives the greatest reduction needs an additional hypothesis. The reduction formula above does not require that assertion. For a counterexample, use the linear kernel with and target . With no informative previous observations, . This can also be arranged with by observing first at , which has zero covariance with all latent values. Observing at the target gives reduction , whereas observing at gives
Indeed this Gaussian process is for a standard normal random variable , so the measurement at observes the same unknown coefficient with a larger signal relative to its noise.
One sufficient extra condition is that the existing latent posterior variance be maximal at among all candidate inputs. To prove it, let and . The Cauchy-Schwarz inequality gives . If , then
because is increasing. Thus target sampling is optimal under this extra condition, but it is not implied by the paper's arbitrary-covariance function assumptions.
Additional observations weakly decrease the posterior variance at every fixed prediction input. Apply part (ii) repeatedly to add observations one at a time. An observation whose remaining conditional covariance with is nonzero strictly decreases that posterior variance; one with zero remaining conditional covariance leaves it unchanged. Large observation noise reduces the information gained through the denominator .
For fixed Gaussian process hyperparameters and a Gaussian likelihood, the posterior variance depends on the observation locations and noise level, but not on the observed numerical outputs. Hence the reduction holds for each possible new output, rather than only on average. The predictive variance of a future noisy observation also decreases by the same amount because its extra independent noise variance is fixed. Refitting the hyperparameters after adding data can change the covariance function; the monotonicity statement compares conditioning under one fixed model.
The Schur complement can be written
because the first term is a conditional variance and . Therefore and the squared term is nonnegative, so
Equality holds exactly when the remaining conditional covariance is zero. In particular, the reduction need not be strict even at a previously unseen input.
Use only the first observations to define
Thus the full noisy covariance matrix is and the full cross-covariance vector is . Set
The scalar is the Schur complement of . Substituting the block inverse into the Gaussian process regression posterior variance gives
Consequently,
and
For interpretation, after the first observations the remaining covariance between and is , while the conditional variance of the new noisy observation is . Thus the reduction is the squared remaining covariance divided by the remaining observation variance.
The conditional multivariate normal distribution gives
where
For a direct justification, form . It is jointly Gaussian with , has zero mean, and satisfies
The independence of uncorrelated jointly normal variables therefore makes independent of . Conditional on , only the displayed affine shift remains, proving the result. This is the latent-function posterior variance; a new observation with fresh independent noise has predictive variance .
Condition throughout on the training inputs and . Interpret here as the fixed covariance function and noise hyperparameters. Assume observation errors are independent standard normal random variables, independent of the Gaussian process , and write
The independent-error assumption is the standard Gaussian process regression interpretation of the observation model; the single-observation marginal noise law alone would not specify cross-observation noise correlations. We take so that is a positive-definite matrix. More generally the inverse formulas below apply whenever the indicated noisy covariance matrices are invertible.
The Gaussian process prior makes a Gaussian random vector. Adding the independent Gaussian noise vector to preserves a multivariate normal distribution. Its mean is zero, its observation covariance matrix is , and because the noise is independent of . Therefore
A degenerate latent variance is allowed; the displayed multivariate normal distribution then denotes a Gaussian law even if it has no full-dimensional probability density.
A recurrent neural network processes positions through a recurrence such as . Computing requires , so the standard computation is sequential across sequence positions, although different examples in a batch can be processed in parallel. Information from position to must pass through recurrent steps, and backpropagated gradients involve products of that many step Jacobian matrices. This creates vanishing gradients or exploding gradients and can make long-range dependencies difficult to learn. Gating can mitigate these difficulties while retaining the recurrent dependence.
In a transformer block, all queries, keys, and values for an available input sequence can be formed in parallel. The self-attention weights are computed by matrix multiplication, so positions within one layer can be processed together; layers themselves remain sequential. With unrestricted self-attention, one position can directly receive information from any other position in one layer, giving a constant-length route across distant positions rather than one proportional to their separation. With causal self-attention, there is the same direct route to any permitted earlier position.
A transformer model supports parallel training across known sequence positions and short routes for long-range dependencies. Even a causal model can compute all training positions together when the entire training sequence is supplied and the attention mask excludes future positions. However, ordinary autoregressive generation still generates one new token at a time. Full self-attention also constructs pairwise scores, with quadratic time and storage in sequence length for the direct implementation; a standard recurrent neural network has linear sequential work in sequence length at fixed hidden width. Neither architecture guarantees that a useful long-range dependency will actually be learned.
A transformer block has two principal sublayers: multi-head attention and a positionwise feedforward neural network. Each sublayer has a residual connection and layer normalization. To make the components explicit, let contain the representations of sequence positions. For head , learned projections form
The scaled dot-product attention output is
where the softmax function acts along each row and is an optional attention mask. For causal self-attention, disallowed future positions receive before the softmax function, giving zero attention weights. In self-attention, queries, keys, and values all come from the same input sequence. The multi-head attention output concatenates the heads and projects them:
A standard post-normalization transformer block can then be written
Here layer normalization normalizes the feature coordinates of each position separately, within each example, and applies learned scales and shifts. The positionwise feedforward neural network applies the same two affine maps and activation function at every position, for example
Thus attention mixes information between positions, while the feedforward sublayer transforms each position's features. The residual connections support propagation through a deep stack, and dropout is often used during training. Pre-normalization variants move layer normalization before the sublayer. A model's positional encoding supplies order information to the sequence representation or attention mechanism; a decoder in an encoder-decoder model also has a cross-attention sublayer between its self-attention and feedforward neural network sublayers.
Evaluating the negative functional gradient at yields
Thus the evaluation vector follows
This is exactly the limiting prediction dynamics in part (ii). If , subtracting the two systems gives with zero initial value, so for every .
An infinitely wide neural network in this regime follows the same training-prediction dynamics as kernel regression with its neural tangent kernel and matching initial predictions. For this squared-error loss function, the relevant algorithm is gradient flow in the Reproducing kernel Hilbert space. Matching predictions at arbitrary additional inputs also requires the corresponding cross-input neural tangent kernel to freeze and the initial predictions there to match. A finite-width neural network can change its feature map during training, so the equivalence depends on the assumptions rather than holding for every architecture and parameterization.
For in the Reproducing kernel Hilbert space , take the directional derivative:
The reproducing property says . Hence
Each kernel section belongs to the Reproducing kernel Hilbert space, so this finite sum is an admissible element. Uniqueness of a vector with these inner products gives the functional gradient
Under the stated infinite-width scaling and freezing assumptions, the deterministic limiting kernel matrix has entries
For a sequence of finite-width networks, this expectation is interpreted in the stated width limit. The resulting neural tangent kernel is a positive-semidefinite kernel, since every finite coefficient vector gives .
The prediction gradient flow is linear:
The matrix exponential solves the constant-coefficient system because the residual satisfies . Each positive-eigenvalue component decays exponentially; a component in the null space of stays at its initial residual. Thus for every initial prediction if is a positive-definite matrix. The limiting kernel matrix being deterministic does not require to be deterministic.
Let be the Jacobian matrix with row . Differentiating the squared-error loss function gives
The parameter gradient flow is consequently . The chain rule then gives
where, explicitly,
It is a Gram matrix and therefore a positive semidefinite matrix: . These derivatives are understood wherever the network is differentiable; for piecewise differentiable activations the same calculation applies along differentiable portions of the trajectory.
Reciprocation reverses the order of positive numbers, so
This is the pullback of the positive-semidefinite kernel from part (i) by . Explicitly, the feature map into the same L2 space is
Hence is a positive-semidefinite kernel, with for every allowed input.
Choose two inputs . Their proposed kernel matrix is
The coefficient vector gives
Thus is not a positive semidefinite matrix, and
Equivalently, its determinant is negative.
Take the real L2 space and the feature map
Each indicator function belongs to because its squared L2 norm is . Their inner product is the length of the overlapping intervals:
This realizes as a positive-semidefinite kernel; it is also the Brownian covariance kernel.

There are unlisted articles, also show them or only show them.