ClickHouse/dbms/include/DB/AggregateFunctions/AggregateFunctionsStatistics.h

#pragma once

#include <DB/IO/WriteHelpers.h>
#include <DB/IO/ReadHelpers.h>
#include <DB/DataTypes/DataTypesNumberFixed.h>
#include <DB/AggregateFunctions/IUnaryAggregateFunction.h>
#include <DB/AggregateFunctions/IBinaryAggregateFunction.h>
#include <DB/Columns/ColumnVector.h>

#include <cmath>

namespace DB
{

namespace
{

/// Эта функция возвращает true если оба значения велики и сравнимы.
/// Она употребляется для вычисления среднего значения путём слияния двух источников.
/// Ибо если размеры обоих источников велики и сравнимы, то надо применить особенную
///	формулу гарантирующую больше стабильности.
bool areComparable(UInt64 a, UInt64 b)
{
	const Float64 sensitivity = 0.001;
	const UInt64 threshold = 10000;

	if ((a == 0) || (b == 0))
		return false;

	auto res = std::minmax(a, b);
	return (((1 - static_cast<Float64>(res.first) / res.second) < sensitivity) && (res.first > threshold));
}

}

/** Статистические аггрегатные функции:
  * varSamp - выборочная дисперсия
  * stddevSamp - среднее выборочное квадратичное отклонение
  * varPop - дисперсия
  * stddevPop - среднее квадратичное отклонение
  * covarSamp - выборочная ковариация
  * covarPop - ковариация
  * corr - корреляция
  */

/** Параллельный и инкрементальный алгоритм для вычисления дисперсии.
  * Источник: "Updating formulae and a pairwise algorithm for computing sample variances"
  * (Chan et al., Stanford University, 12.1979)
  */
template<typename T, typename Op>
class AggregateFunctionVarianceData
{
public:
	AggregateFunctionVarianceData() = default;

	void update(const IColumn & column, size_t row_num)
	{
		T received = static_cast<const ColumnVector<T> &>(column).getData()[row_num];
		Float64 val = static_cast<Float64>(received);
		Float64 delta = val - mean;

		++count;
		mean += delta / count;
		m2 += delta * (val - mean);
	}

	void mergeWith(const AggregateFunctionVarianceData & source)
	{
		UInt64 total_count = count + source.count;
		if (total_count == 0)
			return;

		Float64 factor = static_cast<Float64>(count * source.count) / total_count;
		Float64 delta = mean - source.mean;

		if (areComparable(count, source.count))
			mean = (source.count * source.mean + count * mean) / total_count;
		else
			mean = source.mean + delta * (static_cast<Float64>(count) / total_count);

		m2 += source.m2 + delta * delta * factor;
		count = total_count;
	}

	void serialize(WriteBuffer & buf) const
	{
		writeVarUInt(count, buf);
		writeBinary(mean, buf);
		writeBinary(m2, buf);
	}

	void deserialize(ReadBuffer & buf)
	{
		readVarUInt(count, buf);
		readBinary(mean, buf);
		readBinary(m2, buf);
	}

	void publish(IColumn & to) const
	{
		static_cast<ColumnFloat64 &>(to).getData().push_back(Op::apply(m2, count));
	}

private:
	UInt64 count = 0;
	Float64 mean = 0.0;
	Float64 m2 = 0.0;
};

/** Основной код для реализации функций varSamp, stddevSamp, varPop, stddevPop.
  */
template<typename T, typename Op>
class AggregateFunctionVariance final
	: public IUnaryAggregateFunction<AggregateFunctionVarianceData<T, Op>,
		AggregateFunctionVariance<T, Op> >
{
public:
	String getName() const override { return Op::name; }

	DataTypePtr getReturnType() const override
	{
		return new DataTypeFloat64;
	}

	void setArgument(const DataTypePtr & argument) override
	{
		if (!argument->behavesAsNumber())
			throw Exception("Illegal type " + argument->getName() + " of argument for aggregate function " + getName(),
				ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT);
	}

	void addOne(AggregateDataPtr place, const IColumn & column, size_t row_num) const
	{
		this->data(place).update(column, row_num);
	}

	void merge(AggregateDataPtr place, ConstAggregateDataPtr rhs) const override
	{
		this->data(place).mergeWith(this->data(rhs));
	}

	void serialize(ConstAggregateDataPtr place, WriteBuffer & buf) const override
	{
		this->data(place).serialize(buf);
	}

	void deserializeMerge(AggregateDataPtr place, ReadBuffer & buf) const override
	{
		AggregateFunctionVarianceData<T, Op> source;
		source.deserialize(buf);

		this->data(place).mergeWith(source);
	}

	void insertResultInto(ConstAggregateDataPtr place, IColumn & to) const override
	{
		this->data(place).publish(to);
	}
};

namespace
{

/** Реализации функции varSamp.
  */
struct VarSampImpl
{
	static constexpr auto name = "varSamp";

	static inline Float64 apply(Float64 m2, UInt64 count)
	{
		if (count < 2)
			return std::numeric_limits<Float64>::infinity();
		else
			return m2 / (count - 1);
	}
};

/** Реализация функции stddevSamp.
  */
struct StdDevSampImpl
{
	static constexpr auto name = "stddevSamp";

	static inline Float64 apply(Float64 m2, UInt64 count)
	{
		return sqrt(VarSampImpl::apply(m2, count));
	}
};

/** Реализация функции varPop.
  */
struct VarPopImpl
{
	static constexpr auto name = "varPop";

	static inline Float64 apply(Float64 m2, UInt64 count)
	{
		if (count == 0)
			return std::numeric_limits<Float64>::infinity();
		else if (count == 1)
			return 0.0;
		else
			return m2 / count;
	}
};

/** Реализация функции stddevPop.
  */
struct StdDevPopImpl
{
	static constexpr auto name = "stddevPop";

	static inline Float64 apply(Float64 m2, UInt64 count)
	{
		return sqrt(VarPopImpl::apply(m2, count));
	}
};

}

/** Если флаг compute_marginal_moments установлен, этот класс предоставялет наследнику
  * CovarianceData поддержку маргинальных моментов для вычисления корреляции.
  */
template<bool compute_marginal_moments>
class BaseCovarianceData
{
protected:
	void incrementMarginalMoments(Float64 left_incr, Float64 right_incr) {}
	void mergeWith(const BaseCovarianceData & source) {}
	void serialize(WriteBuffer & buf) const {}
	void deserialize(const ReadBuffer & buf) {}
};

template<>
class BaseCovarianceData<true>
{
protected:
	void incrementMarginalMoments(Float64 left_incr, Float64 right_incr)
	{
		left_m2 += left_incr;
		right_m2 += right_incr;
	}

	void mergeWith(const BaseCovarianceData & source)
	{
		left_m2 += source.left_m2;
		right_m2 += source.right_m2;
	}

	void serialize(WriteBuffer & buf) const
	{
		writeBinary(left_m2, buf);
		writeBinary(right_m2, buf);
	}

	void deserialize(ReadBuffer & buf)
	{
		readBinary(left_m2, buf);
		readBinary(right_m2, buf);
	}

protected:
	Float64 left_m2 = 0.0;
	Float64 right_m2 = 0.0;
};

/** Параллельный и инкрементальный алгоритм для вычисления ковариации.
  * Источник: "Numerically Stable, Single-Pass, Parallel Statistics Algorithms"
  * (J. Bennett et al., Sandia National Laboratories,
  *  2009 IEEE International Conference on Cluster Computing)
  */
template<typename T, typename U, typename Op, bool compute_marginal_moments>
class CovarianceData : public BaseCovarianceData<compute_marginal_moments>
{
private:
	using Base = BaseCovarianceData<compute_marginal_moments>;

public:
	void update(const IColumn & column_left, const IColumn & column_right, size_t row_num)
	{
		T left_received = static_cast<const ColumnVector<T> &>(column_left).getData()[row_num];
		Float64 left_val = static_cast<Float64>(left_received);
		Float64 left_delta = left_val - left_mean;

		U right_received = static_cast<const ColumnVector<U> &>(column_right).getData()[row_num];
		Float64 right_val = static_cast<Float64>(right_received);
		Float64 right_delta = right_val - right_mean;

		Float64 old_right_mean = right_mean;

		++count;

		left_mean +=  left_delta / count;
		right_mean += right_delta / count;
		co_moment += (left_val - left_mean) * (right_val - old_right_mean);

		/// Обновить маргинальные моменты, если они есть.
		if (compute_marginal_moments)
		{
			Float64 left_incr = left_delta * (left_val - left_mean);
			Float64 right_incr = right_delta * (right_val - right_mean);
			Base::incrementMarginalMoments(left_incr, right_incr);
		}
	}

	void mergeWith(const CovarianceData & source)
	{
		UInt64 total_count = count + source.count;
		if (total_count == 0)
			return;

		Float64 factor = static_cast<Float64>(count * source.count) / total_count;
		Float64 left_delta = left_mean - source.left_mean;
		Float64 right_delta = right_mean - source.right_mean;

		if (areComparable(count, source.count))
		{
			left_mean = (source.count * source.left_mean + count * left_mean) / total_count;
			right_mean = (source.count * source.right_mean + count * right_mean) / total_count;
		}
		else
		{
			left_mean = source.left_mean + left_delta * (static_cast<Float64>(count) / total_count);
			right_mean = source.right_mean + right_delta * (static_cast<Float64>(count) / total_count);
		}

		co_moment += source.co_moment + left_delta * right_delta * factor;
		count = total_count;

		/// Обновить маргинальные моменты, если они есть.
		if (compute_marginal_moments)
		{
			Float64 left_incr = left_delta * left_delta * factor;
			Float64 right_incr = right_delta * right_delta * factor;
			Base::mergeWith(source);
			Base::incrementMarginalMoments(left_incr, right_incr);
		}
	}

	void serialize(WriteBuffer & buf) const
	{
		writeVarUInt(count, buf);
		writeBinary(left_mean, buf);
		writeBinary(right_mean, buf);
		writeBinary(co_moment, buf);
		Base::serialize(buf);
	}

	void deserialize(ReadBuffer & buf)
	{
		readVarUInt(count, buf);
		readBinary(left_mean, buf);
		readBinary(right_mean, buf);
		readBinary(co_moment, buf);
		Base::deserialize(buf);
	}

	template<bool compute = compute_marginal_moments>
	void publish(IColumn & to, typename std::enable_if<compute>::type * = nullptr) const
	{
		static_cast<ColumnFloat64 &>(to).getData().push_back(Op::apply(co_moment, Base::left_m2, Base::right_m2, count));
	}

	template<bool compute = compute_marginal_moments>
	void publish(IColumn & to, typename std::enable_if<!compute>::type * = nullptr) const
	{
		static_cast<ColumnFloat64 &>(to).getData().push_back(Op::apply(co_moment, count));
	}

private:
	UInt64 count = 0;
	Float64 left_mean = 0.0;
	Float64 right_mean = 0.0;
	Float64 co_moment = 0.0;
};

template<typename T, typename U, typename Op, bool compute_marginal_moments = false>
class AggregateFunctionCovariance final
	: public IBinaryAggregateFunction<
		CovarianceData<T, U, Op, compute_marginal_moments>,
		AggregateFunctionCovariance<T, U, Op, compute_marginal_moments> >
{
public:
	String getName() const override { return Op::name; }

	DataTypePtr getReturnType() const override
	{
		return new DataTypeFloat64;
	}

	void setArgumentsImpl(const DataTypes & arguments)
	{
		if (!arguments[0]->behavesAsNumber())
			throw Exception("Illegal type " + arguments[0]->getName() + " of first argument to function " + getName(),
				ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT);

		if (!arguments[1]->behavesAsNumber())
			throw Exception("Illegal type " + arguments[1]->getName() + " of second argument to function " + getName(),
				ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT);
	}

	void addOne(AggregateDataPtr place, const IColumn & column_left, const IColumn & column_right, size_t row_num) const
	{
		this->data(place).update(column_left, column_right, row_num);
	}

	void merge(AggregateDataPtr place, ConstAggregateDataPtr rhs) const override
	{
		this->data(place).mergeWith(this->data(rhs));
	}

	void serialize(ConstAggregateDataPtr place, WriteBuffer & buf) const override
	{
		this->data(place).serialize(buf);
	}

	void deserializeMerge(AggregateDataPtr place, ReadBuffer & buf) const override
	{
		CovarianceData<T, U, Op, compute_marginal_moments> source;
		source.deserialize(buf);
		this->data(place).mergeWith(source);
	}

	void insertResultInto(ConstAggregateDataPtr place, IColumn & to) const override
	{
		this->data(place).publish(to);
	}
};

namespace
{

/** Реализация функции covarSamp.
  */
struct CovarSampImpl
{
	static constexpr auto name = "covarSamp";

	static inline Float64 apply(Float64 co_moment, UInt64 count)
	{
		if (count < 2)
			return std::numeric_limits<Float64>::infinity();
		else
			return co_moment / (count - 1);
	}
};

/** Реализация функции covarPop.
  */
struct CovarPopImpl
{
	static constexpr auto name = "covarPop";

	static inline Float64 apply(Float64 co_moment, UInt64 count)
	{
		if (count == 0)
			return std::numeric_limits<Float64>::infinity();
		else if (count == 1)
			return 0.0;
		else
			return co_moment / count;
	}
};

/** Реализация функции corr.
  */
struct CorrImpl
{
	static constexpr auto name = "corr";

	static inline Float64 apply(Float64 co_moment, Float64 left_m2, Float64 right_m2, UInt64 count)
	{
		if (count < 2)
			return std::numeric_limits<Float64>::infinity();
		else
			return co_moment / sqrt(left_m2 * right_m2);
	}
};

}

template<typename T>
using AggregateFunctionVarSamp = AggregateFunctionVariance<T, VarSampImpl>;

template<typename T>
using AggregateFunctionStdDevSamp = AggregateFunctionVariance<T, StdDevSampImpl>;

template<typename T>
using AggregateFunctionVarPop = AggregateFunctionVariance<T, VarPopImpl>;

template<typename T>
using AggregateFunctionStdDevPop = AggregateFunctionVariance<T, StdDevPopImpl>;

template<typename T, typename U>
using AggregateFunctionCovarSamp = AggregateFunctionCovariance<T, U, CovarSampImpl>;

template<typename T, typename U>
using AggregateFunctionCovarPop = AggregateFunctionCovariance<T, U, CovarPopImpl>;

template<typename T, typename U>
using AggregateFunctionCorr = AggregateFunctionCovariance<T, U, CorrImpl, true>;

}