implicit-predicates-gen 0.0.1

//! The error bound of the floating-point filter of a predicate.
//!
//! The filter evaluates the difference form in `f64` and accepts the sign of
//! the result when its magnitude exceeds a bound on the rounding error. The
//! bound is derived here, node by node, in exact arithmetic.

use crate::differences::{DiffGraph, DiffNode};
use crate::dyadic::Dyadic;
use crate::error::Error;
use crate::graph::NodeId;
use crate::parse::Position;

/// The exponent of the limit on the magnitude of every computed node.
const OVERFLOW_EXPONENT: i64 = 1022;

/// The highest degree the filter accepts. Every `c` is at least 1, so a
/// node of higher degree exceeds `2^1022` for every `D > 1`.
pub(crate) const MAX_DEGREE: u32 = 1022;

/// The most bits that the exact coefficients of all the nodes may take
/// together, by the estimate of [`coefficient_bits`].
pub(crate) const MAX_COEFFICIENT_BITS: u64 = 1 << 26;

/// The bounds of one node of degree `g`, in terms of `D` and `E`:
/// `|true value| <= m * D^g`, `|computed value| <= c * D^g` and
/// `|computed - true| <= (alpha * D + beta * E) * D^(g - 1)`.
#[derive(Clone, Debug)]
pub(crate) struct NodeBound {
    pub(crate) degree: u32,
    pub(crate) m: Dyadic,
    pub(crate) c: Dyadic,
    pub(crate) alpha: Dyadic,
    pub(crate) beta: Dyadic,
}

/// The exact error bound of the `f64` evaluation of a difference form.
///
/// For a call, let `delta` be the largest magnitude of the computed leaves
/// (the differences of the coordinates of the approximations of the points),
/// `E` the sum of the two largest errors of the approximations, and
/// `D = delta + E`. The `f64` value of the form differs from the exact value
/// at the exact points by at most `(epsilon * D + kappa * E) * D^(k - 1)`,
/// where `k` is the degree, as long as no product underflows; see
/// [`FilterBound::constants`] for the range in which underflow and overflow
/// are accounted for.
#[derive(Clone, Debug)]
pub struct FilterBound {
    /// The degree `k` of the form in the differences.
    pub degree: u32,
    /// `epsilon / u`, with `u = 2^-53`, as an exact fraction.
    pub epsilon_in_u: String,
    /// `kappa` as an exact fraction.
    pub kappa_exact: String,
    /// The number of products `N` in the form; a product used several times
    /// is counted once.
    pub products: u32,
    nodes: Vec<NodeBound>,
    root: NodeId,
    d_min_exponent: i64,
    d_max_exponent: i64,
}

/// The constants of the filter of a predicate, as `f64`.
///
/// The filter decides when `d_min <= D <= d_max` and
/// `|value| > (epsilon * D + kappa * E) * D^(degree - 1)`.
#[derive(Clone, Copy, Debug, PartialEq)]
pub struct FilterConstants {
    /// The degree `k` of the form.
    pub degree: u32,
    /// The coefficient of `D^k`, rounded up.
    pub epsilon: f64,
    /// The coefficient of `E * D^(k - 1)`, rounded up.
    pub kappa: f64,
    /// The smallest `D` for which the bound covers underflow: a power of two
    /// below 1.
    pub d_min: f64,
    /// The largest `D` for which no node overflows: a power of two above 1.
    pub d_max: f64,
}

/// Derives the error bound of the `f64` evaluation of `graph`.
///
/// With `u = 2^-53`, every node of degree `g` gets coefficients `m`, `c`,
/// `alpha` and `beta` such that `|true value| <= m * D^g`,
/// `|computed value| <= c * D^g` and
/// `|computed - true| <= (alpha * D + beta * E) * D^(g - 1)`:
///
/// - a leaf has `g = 1`, `m = 1 + u`, `c = 1`, `alpha = u`, `beta = 1`;
/// - a negation is exact and keeps the coefficients of its operand;
/// - the sum or difference of `x` and `y`, of equal degree, has
///   `m = m_x + m_y`, `c = (c_x + c_y)(1 + u)`,
///   `alpha = alpha_x + alpha_y + u (c_x + c_y)` and `beta = beta_x + beta_y`;
/// - the product of `x` and `y` has `g = g_x + g_y`, `m = m_x m_y`,
///   `c = c_x c_y (1 + u)`, `alpha = c_x alpha_y + m_y alpha_x + u c_x c_y`
///   and `beta = c_x beta_y + m_y beta_x`.
///
/// The root gives `epsilon = alpha` and `kappa = beta`. Every node is
/// analysed once, in the order of the graph, so a shared subexpression counts
/// as the single computation it is.
///
/// # Errors
///
/// Returns an error, at `position` (the name of the predicate) in `file`,
/// when the degree is so high that no `D > 1` keeps every computed node
/// below `2^1022`, or that no `D < 1` keeps the bound clear of underflow (see
/// [`FilterBound::constants`]), and when the exact coefficients would take
/// more than `2^26` bits in all. The coefficients of a sum are about 53
/// bits longer than those of its operands, so a sum written as one chain of
/// more than about 1,560 terms reaches that limit; the same terms summed in a
/// balanced tree, with `let`, need far fewer bits. The size is estimated
/// node by node before any coefficient is computed, so a definition that
/// is too large fails quickly.
pub fn filter_bound(
    graph: &DiffGraph,
    file: &str,
    position: Position,
) -> Result<FilterBound, Error> {
    if let Some(&degree) = graph.degrees.iter().filter(|&&g| g > MAX_DEGREE).max() {
        return Err(Error::at(file, position, overflow_message(degree)));
    }
    if coefficient_bits(graph) > MAX_COEFFICIENT_BITS {
        return Err(Error::at(
            file,
            position,
            format!(
                "the expression is too large for the error bound of the filter: its exact \
                 coefficients would take more than 2^{} bits in all; a long sum written as \
                 one chain is the usual cause, and the same terms summed in a balanced \
                 tree with `let` need far fewer",
                MAX_COEFFICIENT_BITS.ilog2()
            ),
        ));
    }
    let u = Dyadic::pow2(-53);
    let one = Dyadic::integer(1);
    let one_plus_u = &one + &u;
    let mut nodes: Vec<NodeBound> = Vec::with_capacity(graph.nodes.len());
    for (id, node) in graph.nodes.iter().enumerate() {
        let bound = match *node {
            DiffNode::Leaf { .. } => NodeBound {
                degree: 1,
                m: one_plus_u.clone(),
                c: one.clone(),
                alpha: u.clone(),
                beta: one.clone(),
            },
            DiffNode::Neg(x) => nodes[x].clone(),
            DiffNode::Add(x, y) | DiffNode::Sub(x, y) => {
                let (x, y) = (&nodes[x], &nodes[y]);
                let c_sum = &x.c + &y.c;
                NodeBound {
                    degree: x.degree,
                    m: &x.m + &y.m,
                    c: &c_sum * &one_plus_u,
                    alpha: &(&x.alpha + &y.alpha) + &(&u * &c_sum),
                    beta: &x.beta + &y.beta,
                }
            }
            DiffNode::Mul(x, y) => {
                let (x, y) = (&nodes[x], &nodes[y]);
                let c_product = &x.c * &y.c;
                NodeBound {
                    degree: x.degree + y.degree,
                    m: &x.m * &y.m,
                    c: &c_product * &one_plus_u,
                    alpha: &(&(&x.c * &y.alpha) + &(&y.m * &x.alpha)) + &(&u * &c_product),
                    beta: &(&x.c * &y.beta) + &(&y.m * &x.beta),
                }
            }
        };
        debug_assert_eq!(bound.degree, graph.degrees[id]);
        nodes.push(bound);
    }
    let products = graph
        .nodes
        .iter()
        .filter(|n| matches!(n, DiffNode::Mul(..)))
        .count();
    let root = &nodes[graph.root];
    let degree = root.degree;
    // Every node takes at least 55 bits, so the budget allows fewer than
    // 2^19 nodes.
    let products = u32::try_from(products).expect("fewer than 2^32 products");
    let d_max_exponent = d_max_exponent(&nodes)
        .map_err(|degree| Error::at(file, position, overflow_message(degree)))?;
    let d_min_exponent = d_min_exponent(&nodes, root, products).ok_or_else(|| {
        Error::at(
            file,
            position,
            format!(
                "the degree {degree} is too high for the floating-point filter: \
                 no scale of the inputs below 1 keeps its error bound clear of underflow"
            ),
        )
    })?;
    Ok(FilterBound {
        degree,
        epsilon_in_u: (&root.alpha * &Dyadic::pow2(53)).fraction(),
        kappa_exact: root.beta.fraction(),
        products,
        nodes,
        root: graph.root,
        d_min_exponent,
        d_max_exponent,
    })
}

/// An upper bound on the bits of all the exact coefficients of `graph`
/// together, found without computing them, saturated at `u64::MAX`.
///
/// Every coefficient of a node is `P(u)` for a polynomial `P` with
/// nonnegative integer coefficients, so it is an integer times `2^(-53 d)`,
/// with `d` the degree of `P`, of magnitude at most `P(1)`: its mantissa
/// has at most `53 d + log2 P(1) + 1` bits. Following the rules of
/// [`filter_bound`], that is at most 55 bits for a leaf, the operand's
/// for a negation, 55 more than the larger operand's for a sum or
/// difference, and 55 more than the two operands' together for a product.
/// The result is the sum over the nodes.
fn coefficient_bits(graph: &DiffGraph) -> u64 {
    let mut bits: Vec<u64> = Vec::with_capacity(graph.nodes.len());
    for node in &graph.nodes {
        bits.push(match *node {
            DiffNode::Leaf { .. } => 55,
            DiffNode::Neg(x) => bits[x],
            DiffNode::Add(x, y) | DiffNode::Sub(x, y) => bits[x].max(bits[y]).saturating_add(55),
            DiffNode::Mul(x, y) => bits[x].saturating_add(bits[y]).saturating_add(55),
        });
    }
    bits.iter().fold(0u64, |total, &b| total.saturating_add(b))
}

fn overflow_message(degree: u32) -> String {
    format!(
        "the degree {degree} is too high for the floating-point filter: \
         no scale of the inputs above 1 keeps every computed value below 2^{OVERFLOW_EXPONENT}"
    )
}

impl FilterBound {
    /// Whether `n u <= epsilon < n u (1 + 2^-40)`, exactly.
    ///
    /// The exact `epsilon` is slightly above an integer multiple of `u`,
    /// because of the terms in `u^2` and beyond.
    pub fn epsilon_is_u_times(&self, n: u64) -> bool {
        let target = &Dyadic::integer(n) * &Dyadic::pow2(-53);
        close_above(&self.nodes[self.root].alpha, &target)
    }

    /// Whether `n <= kappa < n (1 + 2^-40)`, exactly.
    pub fn kappa_is(&self, n: u64) -> bool {
        close_above(&self.nodes[self.root].beta, &Dyadic::integer(n))
    }

    /// The constants of the filter.
    ///
    /// `epsilon` and `kappa` are the exact values times `1 + 2^-30`, rounded
    /// up; the inflation covers the rounding of the evaluation of the bound
    /// itself and an allowance for products that underflow. `d_min` is the
    /// smallest power of two below 1 for which that allowance covers
    /// underflow and the bound stays a normal number, and `d_max` the
    /// largest above 1 for which no computed node exceeds `2^1022`; both
    /// exist, or [`filter_bound`] would have returned an error.
    pub fn constants(&self) -> FilterConstants {
        let inflate = &Dyadic::integer(1) + &Dyadic::pow2(-30);
        let root = &self.nodes[self.root];
        FilterConstants {
            degree: self.degree,
            epsilon: (&root.alpha * &inflate).round_up(),
            kappa: (&root.beta * &inflate).round_up(),
            d_min: Dyadic::pow2(self.d_min_exponent).round_up(),
            d_max: Dyadic::pow2(self.d_max_exponent).round_up(),
        }
    }
}

/// The exponent of `d_max`: the largest `j >= 1` such that
/// `c_x * 2^(j g_x) <= 2^1022` at every node `x`.
///
/// For `D <= d_max` every computed node then has magnitude at most
/// `c_x * D^(g_x) <= 2^1022` (for `D < 1`, `D^(g_x) <= 1 < d_max^(g_x)`),
/// a quarter of the overflow threshold `2^1024`, so no intermediate value
/// and not the result overflows. An overflow in the evaluation of the bound
/// itself gives an infinite bound, which only makes the filter abstain.
///
/// Returns the degree of a node that overflows for every `j >= 1` if there is
/// no such `j`.
fn d_max_exponent(nodes: &[NodeBound]) -> Result<i64, u32> {
    // `c_x * 2^(j g) <= 2^1022` iff `j g <= 1022 - s` with `s` the
    // smallest integer such that `c_x <= 2^s`.
    let (j, degree) = nodes
        .iter()
        .map(|n| {
            let j = (OVERFLOW_EXPONENT - n.c.ceil_log2()).div_euclid(i64::from(n.degree));
            (j, n.degree)
        })
        .min()
        .expect("a graph has a root");
    if j >= 1 { Ok(j) } else { Err(degree) }
}

/// The exponent of `d_min`: the smallest `j <= -1` such that `P = 2^j`
/// satisfies
///
/// 1. `epsilon * P^k * 2^-32 >= N * k * c_max * 2^-1074`, with `N` the
///    number of products and `c_max` the largest `c` of any node, and
/// 2. `epsilon * P^k >= 2^-1021`.
///
/// Both then hold for every larger `j`.
///
/// Why this is sound, for `D >= d_min` and with `B` the rounding-error
/// bound `(epsilon * D + kappa * E) * D^(k - 1)`:
///
/// - The node rules assume that every rounding error is at most `u` times
///   the computed magnitude. A sum or difference whose result is
///   subnormal is exact, but a product that underflows can err by up to
///   `2^-1075` in absolute terms; allow `2^-1074` per product.
/// - An absolute error at a product `x` reaches the root multiplied by
///   the derivative of the computation with respect to `x`: a sum over
///   the paths from `x` to the root of the products of the other factors
///   met on the way, each with its `(1 + u)` rounding. Every factor is
///   bounded by its `c * D^g`, so the derivative is at most
///   `p'(c_x) * D^(k - g_x)`, where `p(t)` is `c_root` computed by the
///   rules with `c_x` replaced by `t`. That polynomial has nonnegative
///   coefficients and degree at most `k / g_x`, since degrees add along
///   products, so `c_x p'(c_x) <= (k / g_x) p(c_x)`; with `c_x >= 1` and
///   `g_x >= 2` the derivative is at most `k * c_max * D^(k - g_x)`.
///   Terms of second order in these errors are products of two such
///   errors, each at most `2^-1074` times factors bounded as above, so they
///   are smaller than the first-order terms by a factor of about `2^-1000`
///   and are absorbed by the slack of `2^-32` of the bound below.
/// - The total underflow error is therefore
///   `U <= N * k * c_max * 2^-1074 * max(1, D^k)`. For `d_min <= D <= 1`,
///   condition 1 gives `U <= 2^-32 * epsilon * d_min^k
///   <= 2^-32 * epsilon * D^k`; for `D > 1` it gives the same, because
///   condition 1 holds with `d_min^k < 1`. So `U <= 2^-32 * B`, and the
///   value of the form errs by at most `(1 + 2^-32) * B`.
/// - The emitted constants are at least `(1 + 2^-30)` times the exact
///   ones, so the exact value of the emitted bound is at least
///   `(1 + 2^-30) * B`. Its `f64` evaluation, `d = delta + e`, then
///   `(EPSILON * d + KAPPA * e) * d * ... * d` with `k - 1` products by
///   `d`, takes `k + 3` operations, each of which can round down by a
///   factor `1 - u`. The rounding of `d` enters the bound once for each of
///   the `k` times `d` is used, so the computed bound is at least
///   `(1 - u)^(2k + 2)` times its exact value: `k + 3` factors for the
///   operations and `k - 1` more for the uses of `d` beyond the first
///   (`e` is an upper bound on `E`, rounded up by the caller, so it adds
///   none; a caller that rounded it to nearest instead would add one factor
///   for each use of `e` and of `d`, `3k + 3` in all, within `2k + 8` only
///   for `k <= 5`). That holds while no partial result is subnormal:
///   condition 2 keeps every partial result built from `EPSILON * d` at least
///   `epsilon * min(1, D)^k >= 2^-1021`, and the term `KAPPA * e`, which
///   may underflow, loses at most `2^-1075`, a fraction `2^-54` of the sum
///   it joins. The constants were derived for up to `2k + 8` such factors,
///   and `(1 + 2^-30)(1 - u)^(2k + 8) >= 1 + 2^-31` for `k <= 1022` (higher
///   degrees are rejected), so the computed bound is at least
///   `(1 + 2^-31) * B`, which exceeds the error `(1 + 2^-32) * B`.
///
/// Returns `None` if no `j <= -1` satisfies both.
fn d_min_exponent(nodes: &[NodeBound], root: &NodeBound, products: u32) -> Option<i64> {
    let epsilon = &root.alpha;
    let c_max = nodes
        .iter()
        .map(|n| &n.c)
        .max()
        .expect("a graph has a root");
    let factor = Dyadic::integer(u64::from(products) * u64::from(root.degree));
    let underflow = &(&factor * c_max) * &Dyadic::pow2(-1074 + 32);
    let normal = Dyadic::pow2(-1021);
    let k = i64::from(root.degree);
    let covers = |j: i64| {
        let scaled = epsilon * &Dyadic::pow2(j * k);
        scaled >= underflow && scaled >= normal
    };
    if !covers(-1) {
        return None;
    }
    let mut j = -1;
    while covers(j - 1) {
        j -= 1;
    }
    Some(j)
}

/// Whether `target <= value < target * (1 + 2^-40)`.
fn close_above(value: &Dyadic, target: &Dyadic) -> bool {
    let limit = target * &(&Dyadic::integer(1) + &Dyadic::pow2(-40));
    target <= value && *value < limit
}

#[cfg(test)]
mod tests;