pyoframe 0.2.0__py3-none-any.whl → 1.0.0a0__py3-none-any.whl

pyoframe/_core.py

@@ -0,0 +1,2575 @@
+"""Defines several core Pyoframe objects including Set, Constraint, Variable, and Expression."""
+
+from __future__ import annotations
+
+import warnings
+from abc import abstractmethod
+from collections.abc import Iterable, Mapping, Sequence
+from typing import TYPE_CHECKING, Literal, Protocol, Union, overload
+
+import pandas as pd
+import polars as pl
+import pyoptinterface as poi
+
+from pyoframe._arithmetic import (
+    _get_dimensions,
+    _simplify_expr_df,
+    add,
+    multiply,
+)
+from pyoframe._constants import (
+    COEF_KEY,
+    CONST_TERM,
+    CONSTRAINT_KEY,
+    DUAL_KEY,
+    KEY_TYPE,
+    QUAD_VAR_KEY,
+    RESERVED_COL_KEYS,
+    SOLUTION_KEY,
+    VAR_KEY,
+    Config,
+    ConstraintSense,
+    ObjSense,
+    PyoframeError,
+    UnmatchedStrategy,
+    VType,
+    VTypeValue,
+)
+from pyoframe._model_element import ModelElement, ModelElementWithId
+from pyoframe._utils import (
+    Container,
+    FuncArgs,
+    cast_coef_to_string,
+    concat_dimensions,
+    get_obj_repr,
+    pairwise,
+    parse_inputs_as_iterable,
+    return_new,
+    unwrap_single_values,
+)
+
+if TYPE_CHECKING:  # pragma: no cover
+    from pyoframe._model import Model
+
+
+# TODO consider changing this simply to a type and having a helper "Expression.from(object)"
+class SupportsToExpr(Protocol):
+    """Protocol for any object that can be converted to a Pyoframe [Expression][pyoframe.Expression]."""
+
+    def to_expr(self) -> Expression:
+        """Converts the object to a Pyoframe [Expression][pyoframe.Expression]."""
+        ...
+
+
+class SupportsMath(ModelElement, SupportsToExpr):
+    """Any object that can be converted into an expression."""
+
+    def __init__(self, *args, **kwargs):
+        self._unmatched_strategy = UnmatchedStrategy.UNSET
+        self._allowed_new_dims: list[str] = []
+        super().__init__(*args, **kwargs)
+
+    @abstractmethod
+    def _new(self, data: pl.DataFrame, name: str) -> SupportsMath:
+        """Helper method to create a new instance of the same (or for Variable derivative) class."""
+
+    def _copy_flags(self, other: SupportsMath):
+        """Copies the flags from another SupportsMath object."""
+        self._unmatched_strategy = other._unmatched_strategy
+        self._allowed_new_dims = other._allowed_new_dims.copy()
+
+    def keep_unmatched(self):
+        """Indicates that all rows should be kept during addition or subtraction, even if they are not matched in the other expression."""
+        new = self._new(self.data, name=f"{self.name}.keep_unmatched()")
+        new._copy_flags(self)
+        new._unmatched_strategy = UnmatchedStrategy.KEEP
+        return new
+
+    def drop_unmatched(self):
+        """Indicates that rows that are not matched in the other expression during addition or subtraction should be dropped."""
+        new = self._new(self.data, name=f"{self.name}.drop_unmatched()")
+        new._copy_flags(self)
+        new._unmatched_strategy = UnmatchedStrategy.DROP
+        return new
+
+    def over(self, *dims: str):
+        """Indicates that the expression can be broadcasted over the given dimensions during addition and subtraction."""
+        new = self._new(self.data, name=f"{self.name}.over(…)")
+        new._copy_flags(self)
+        new._allowed_new_dims.extend(dims)
+        return new
+
+    @return_new
+    def rename(self, *args, **kwargs):
+        """Renames one or several of the object's dimensions.
+
+        Takes the same arguments as [`polars.DataFrame.rename`](https://docs.pola.rs/api/python/stable/reference/dataframe/api/polars.DataFrame.rename.html).
+
+        See the [portfolio optimization example](../examples/portfolio_optimization.md) for a usage example.
+
+        Examples:
+            >>> m = pf.Model()
+            >>> m.v = pf.Variable(
+            ...     {"hour": ["00:00", "06:00", "12:00", "18:00"]},
+            ...     {"city": ["Toronto", "Berlin", "Paris"]},
+            ... )
+            >>> m.v
+            <Variable 'v' height=12>
+            ┌───────┬─────────┬──────────────────┐
+            │ hour  ┆ city    ┆ variable         │
+            │ (4)   ┆ (3)     ┆                  │
+            ╞═══════╪═════════╪══════════════════╡
+            │ 00:00 ┆ Toronto ┆ v[00:00,Toronto] │
+            │ 00:00 ┆ Berlin  ┆ v[00:00,Berlin]  │
+            │ 00:00 ┆ Paris   ┆ v[00:00,Paris]   │
+            │ 06:00 ┆ Toronto ┆ v[06:00,Toronto] │
+            │ 06:00 ┆ Berlin  ┆ v[06:00,Berlin]  │
+            │ …     ┆ …       ┆ …                │
+            │ 12:00 ┆ Berlin  ┆ v[12:00,Berlin]  │
+            │ 12:00 ┆ Paris   ┆ v[12:00,Paris]   │
+            │ 18:00 ┆ Toronto ┆ v[18:00,Toronto] │
+            │ 18:00 ┆ Berlin  ┆ v[18:00,Berlin]  │
+            │ 18:00 ┆ Paris   ┆ v[18:00,Paris]   │
+            └───────┴─────────┴──────────────────┘
+
+            >>> m.v.rename({"city": "location"})
+            <Expression height=12 terms=12 type=linear>
+            ┌───────┬──────────┬──────────────────┐
+            │ hour  ┆ location ┆ expression       │
+            │ (4)   ┆ (3)      ┆                  │
+            ╞═══════╪══════════╪══════════════════╡
+            │ 00:00 ┆ Toronto  ┆ v[00:00,Toronto] │
+            │ 00:00 ┆ Berlin   ┆ v[00:00,Berlin]  │
+            │ 00:00 ┆ Paris    ┆ v[00:00,Paris]   │
+            │ 06:00 ┆ Toronto  ┆ v[06:00,Toronto] │
+            │ 06:00 ┆ Berlin   ┆ v[06:00,Berlin]  │
+            │ …     ┆ …        ┆ …                │
+            │ 12:00 ┆ Berlin   ┆ v[12:00,Berlin]  │
+            │ 12:00 ┆ Paris    ┆ v[12:00,Paris]   │
+            │ 18:00 ┆ Toronto  ┆ v[18:00,Toronto] │
+            │ 18:00 ┆ Berlin   ┆ v[18:00,Berlin]  │
+            │ 18:00 ┆ Paris    ┆ v[18:00,Paris]   │
+            └───────┴──────────┴──────────────────┘
+
+        """
+        return self.data.rename(*args, **kwargs)
+
+    @return_new
+    def with_columns(self, *args, **kwargs):
+        """Creates a new object with modified columns.
+
+        Takes the same arguments as [`polars.DataFrame.with_columns`](https://docs.pola.rs/api/python/stable/reference/dataframe/api/polars.DataFrame.with_columns.html).
+
+        !!! warning
+            Only use this function if you know what you're doing. It is not recommended to manually modify the columns
+            within a Pyoframe object.
+        """
+        return self.data.with_columns(*args, **kwargs)
+
+    @return_new
+    def filter(self, *args, **kwargs):
+        """Creates a copy of the object containing only a subset of the original rows.
+
+        Takes the same arguments as [`polars.DataFrame.filter`](https://docs.pola.rs/api/python/stable/reference/dataframe/api/polars.DataFrame.filter.html).
+
+        See Also:
+            [`Expression.pick`][pyoframe.Expression.pick] or [`Variable.pick`][pyoframe.Variable.pick] if you wish to drop the filtered
+            column in the process.
+
+        """
+        return self.data.filter(*args, **kwargs)
+
+    @return_new
+    def pick(self, **kwargs):
+        """Filters elements by the given criteria and then drops the filtered dimensions.
+
+        Examples:
+            >>> m = pf.Model()
+            >>> m.v = pf.Variable(
+            ...     [
+            ...         {"hour": ["00:00", "06:00", "12:00", "18:00"]},
+            ...         {"city": ["Toronto", "Berlin", "Paris"]},
+            ...     ]
+            ... )
+            >>> m.v.pick(hour="06:00")
+            <Expression height=3 terms=3 type=linear>
+            ┌─────────┬──────────────────┐
+            │ city    ┆ expression       │
+            │ (3)     ┆                  │
+            ╞═════════╪══════════════════╡
+            │ Toronto ┆ v[06:00,Toronto] │
+            │ Berlin  ┆ v[06:00,Berlin]  │
+            │ Paris   ┆ v[06:00,Paris]   │
+            └─────────┴──────────────────┘
+            >>> m.v.pick(hour="06:00", city="Toronto")
+            <Expression terms=1 type=linear>
+            v[06:00,Toronto]
+
+        See Also:
+            [`Expression.filter`][pyoframe.Expression.filter] or [`Variable.filter`][pyoframe.Variable.filter] if you don't wish to drop the filtered column.
+        """
+        return self.data.filter(**kwargs).drop(kwargs.keys())
+
+    def _add_allowed_new_dims_to_df(self, df):
+        cols = df.columns
+        df = df.with_columns(*(pl.lit("*").alias(c) for c in self._allowed_new_dims))
+        df = df.select(cols[:-1] + self._allowed_new_dims + [cols[-1]])  # reorder
+        return df
+
+    def add_dim(self, *dims: str):  # pragma: no cover
+        """Deprecated, use [`over`][pyoframe.Expression.over] instead."""
+        warnings.warn(
+            "'add_dim' has been renamed to 'over'. Please use 'over' instead.",
+            DeprecationWarning,
+        )
+        return self.over(*dims)
+
+    @abstractmethod
+    def to_expr(self) -> Expression:
+        """Converts the object to a Pyoframe Expression."""
+        ...
+
+    def sum(self, *args, **kwargs):
+        """Converts the object to an expression (see `.to_expr()`) and then applies [`Expression.sum`][pyoframe.Expression.sum]."""
+        return self.to_expr().sum(*args, **kwargs)
+
+    def sum_by(self, *args, **kwargs):
+        """Converts the object to an expression (see `.to_expr()`) and then applies [`Expression.sum_by`][pyoframe.Expression.sum_by]."""
+        return self.to_expr().sum_by(*args, **kwargs)
+
+    def map(self, *args, **kwargs):
+        """Converts the object to an expression (see `.to_expr()`) and then applies [`Expression.map`][pyoframe.Expression.map]."""
+        return self.to_expr().map(*args, **kwargs)
+
+    def __add__(self, *args, **kwargs):
+        return self.to_expr().__add__(*args, **kwargs)
+
+    def __mul__(self, *args, **kwargs):
+        return self.to_expr().__mul__(*args, **kwargs)
+
+    def __pow__(self, power: int):
+        """Supports squaring expressions.
+
+        Examples:
+            >>> m = pf.Model()
+            >>> m.v = pf.Variable()
+            >>> m.v**2
+            <Expression terms=1 type=quadratic>
+            v * v
+            >>> m.v**3
+            Traceback (most recent call last):
+            ...
+            ValueError: Raising an expressions to **3 is not supported. Expressions can only be squared (**2).
+        """
+        if power == 2:
+            res = self * self
+            res.name = f"({self.name}**2)"
+            return res
+        raise ValueError(
+            f"Raising an expressions to **{power} is not supported. Expressions can only be squared (**2)."
+        )
+
+    def __neg__(self):
+        res = self.to_expr() * -1
+        res.name = f"-{self.name}"
+        res._copy_flags(self)
+        return res
+
+    def __sub__(self, other):
+        """Subtracts a value from this Expression.
+
+        Examples:
+            >>> import polars as pl
+            >>> m = pf.Model()
+            >>> df = pl.DataFrame({"dim1": [1, 2, 3], "value": [1, 2, 3]})
+            >>> m.v = pf.Variable(df["dim1"])
+            >>> m.v - df
+            <Expression height=3 terms=6 type=linear>
+            ┌──────┬────────────┐
+            │ dim1 ┆ expression │
+            │ (3)  ┆            │
+            ╞══════╪════════════╡
+            │ 1    ┆ v[1] -1    │
+            │ 2    ┆ v[2] -2    │
+            │ 3    ┆ v[3] -3    │
+            └──────┴────────────┘
+        """
+        if not isinstance(other, (int, float)):
+            other = other.to_expr()
+        return self.to_expr() + (-other)
+
+    def __rmul__(self, other):
+        return self.to_expr() * other
+
+    def __radd__(self, other):
+        return self.to_expr() + other
+
+    def __truediv__(self, other):
+        """Divides this expression.
+
+        Examples:
+            Support division.
+            >>> m = pf.Model()
+            >>> m.v = Variable({"dim1": [1, 2, 3]})
+            >>> m.v / 2
+            <Expression height=3 terms=3 type=linear>
+            ┌──────┬────────────┐
+            │ dim1 ┆ expression │
+            │ (3)  ┆            │
+            ╞══════╪════════════╡
+            │ 1    ┆ 0.5 v[1]   │
+            │ 2    ┆ 0.5 v[2]   │
+            │ 3    ┆ 0.5 v[3]   │
+            └──────┴────────────┘
+        """
+        return self.to_expr() * (1 / other)
+
+    def __rsub__(self, other):
+        """Supports right subtraction.
+
+        Examples:
+            >>> m = pf.Model()
+            >>> m.v = Variable({"dim1": [1, 2, 3]})
+            >>> 1 - m.v
+            <Expression height=3 terms=6 type=linear>
+            ┌──────┬────────────┐
+            │ dim1 ┆ expression │
+            │ (3)  ┆            │
+            ╞══════╪════════════╡
+            │ 1    ┆ 1 - v[1]   │
+            │ 2    ┆ 1 - v[2]   │
+            │ 3    ┆ 1 - v[3]   │
+            └──────┴────────────┘
+        """
+        return other + (-self.to_expr())
+
+    def __le__(self, other):
+        """Equality constraint.
+
+        Examples:
+            >>> m = pf.Model()
+            >>> m.v = pf.Variable()
+            >>> m.v <= 1
+            <Constraint 'unnamed' terms=2 type=linear>
+            v <= 1
+        """
+        return Constraint(self - other, ConstraintSense.LE)
+
+    def __ge__(self, other):
+        """Equality constraint.
+
+        Examples:
+            >>> m = pf.Model()
+            >>> m.v = pf.Variable()
+            >>> m.v >= 1
+            <Constraint 'unnamed' terms=2 type=linear>
+            v >= 1
+        """
+        return Constraint(self - other, ConstraintSense.GE)
+
+    def __eq__(self, value: object):  # type: ignore
+        """Equality constraint.
+
+        Examples:
+            >>> m = pf.Model()
+            >>> m.v = pf.Variable()
+            >>> m.v == 1
+            <Constraint 'unnamed' terms=2 type=linear>
+            v = 1
+        """
+        return Constraint(self - value, ConstraintSense.EQ)
+
+
+SetTypes = Union[
+    pl.DataFrame,
+    pd.Index,
+    pd.DataFrame,
+    SupportsMath,
+    Mapping[str, Sequence[object]],
+    "Set",
+    "Constraint",
+]
+
+
+class Set(SupportsMath):
+    """A set which can then be used to index variables.
+
+    Examples:
+        >>> pf.Set(x=range(2), y=range(3))
+        <Set 'unnamed' height=6>
+        ┌─────┬─────┐
+        │ x   ┆ y   │
+        │ (2) ┆ (3) │
+        ╞═════╪═════╡
+        │ 0   ┆ 0   │
+        │ 0   ┆ 1   │
+        │ 0   ┆ 2   │
+        │ 1   ┆ 0   │
+        │ 1   ┆ 1   │
+        │ 1   ┆ 2   │
+        └─────┴─────┘
+    """
+
+    def __init__(self, *data: SetTypes | Iterable[SetTypes], **named_data):
+        data_list = list(data)
+        for name, set in named_data.items():
+            data_list.append({name: set})
+        df = self._parse_acceptable_sets(*data_list)
+        if not df.is_empty() and df.is_duplicated().any():
+            raise ValueError("Duplicate rows found in input data.")
+        super().__init__(df, name="unnamed_set")
+
+    def _new(self, data: pl.DataFrame, name: str) -> Set:
+        s = Set(data)
+        s.name = name
+        s._model = self._model
+        return s
+
+    @staticmethod
+    def _parse_acceptable_sets(
+        *over: SetTypes | Iterable[SetTypes],
+    ) -> pl.DataFrame:
+        """Computes the cartesian product of the given sets.
+
+        Examples:
+            >>> import pandas as pd
+            >>> dim1 = pd.Index([1, 2, 3], name="dim1")
+            >>> dim2 = pd.Index(["a", "b"], name="dim1")
+            >>> Set._parse_acceptable_sets([dim1, dim2])
+            Traceback (most recent call last):
+            ...
+            AssertionError: Dimension 'dim1' is not unique.
+            >>> dim2.name = "dim2"
+            >>> Set._parse_acceptable_sets([dim1, dim2])
+            shape: (6, 2)
+            ┌──────┬──────┐
+            │ dim1 ┆ dim2 │
+            │ ---  ┆ ---  │
+            │ i64  ┆ str  │
+            ╞══════╪══════╡
+            │ 1    ┆ a    │
+            │ 1    ┆ b    │
+            │ 2    ┆ a    │
+            │ 2    ┆ b    │
+            │ 3    ┆ a    │
+            │ 3    ┆ b    │
+            └──────┴──────┘
+        """
+        assert len(over) > 0, "At least one set must be provided."
+        over_iter: Iterable[SetTypes] = parse_inputs_as_iterable(*over)
+
+        over_frames: list[pl.DataFrame] = [Set._set_to_polars(set) for set in over_iter]
+
+        over_merged = over_frames[0]
+
+        for df in over_frames[1:]:
+            overlap_dims = set(over_merged.columns) & set(df.columns)
+            assert not overlap_dims, (
+                f"Dimension '{tuple(overlap_dims)[0]}' is not unique."
+            )
+            over_merged = over_merged.join(df, how="cross")
+        return over_merged
+
+    def to_expr(self) -> Expression:
+        """Converts the Set to an Expression equal to 1 for each index.
+
+        Useful when multiplying a Set by an Expression.
+        """
+        return Expression(
+            self.data.with_columns(
+                pl.lit(1).alias(COEF_KEY), pl.lit(CONST_TERM).alias(VAR_KEY)
+            ),
+            name=self.name,
+        )
+
+    def __mul__(self, other):
+        if isinstance(other, Set):
+            overlap_dims = set(self.data.columns) & set(other.data.columns)
+            assert not overlap_dims, (
+                f"Cannot multiply the two sets because dimension '{tuple(overlap_dims)[0]}' is present in both sets."
+            )
+            return Set(self.data, other.data)
+        return super().__mul__(other)
+
+    def __add__(self, other):
+        # TODO replace with bitwise or
+        if isinstance(other, Set):
+            try:
+                return self._new(
+                    pl.concat([self.data, other.data]).unique(
+                        maintain_order=Config.maintain_order
+                    ),
+                    name=f"({self.name} + {other.name})",
+                )
+            except pl.exceptions.ShapeError as e:
+                if "unable to vstack, column names don't match" in str(e):
+                    raise PyoframeError(
+                        f"Failed to add sets '{self.name}' and '{other.name}' because dimensions do not match ({self.dimensions} != {other.dimensions}) "
+                    ) from e
+                raise e  # pragma: no cover
+
+        return super().__add__(other)
+
+    def __repr__(self):
+        header = get_obj_repr(
+            self,
+            "unnamed" if self.name == "unnamed_set" else self.name,
+            height=self.data.height,
+        )
+        data = self._add_shape_to_columns(self.data)
+        data = self._add_allowed_new_dims_to_df(data)
+        with Config.print_polars_config:
+            table = repr(data)
+
+        return header + "\n" + table
+
+    @staticmethod
+    def _set_to_polars(set: SetTypes) -> pl.DataFrame:
+        if isinstance(set, dict):
+            df = pl.DataFrame(set)
+        elif isinstance(set, Constraint):
+            df = set.data.select(set._dimensions_unsafe)
+        elif isinstance(set, SupportsMath):
+            df = (
+                set.to_expr()
+                .data.drop(RESERVED_COL_KEYS, strict=False)
+                .unique(maintain_order=Config.maintain_order)
+            )
+        elif isinstance(set, pd.Index):
+            df = pl.from_pandas(pd.DataFrame(index=set).reset_index())
+        elif isinstance(set, pd.DataFrame):
+            df = pl.from_pandas(set)
+        elif isinstance(set, pl.DataFrame):
+            df = set
+        elif isinstance(set, pl.Series):
+            df = set.to_frame()
+        elif isinstance(set, pd.Series):
+            if not set.name:
+                raise ValueError("Cannot convert an unnamed Pandas Series to a Set.")
+            df = pl.from_pandas(set).to_frame()
+        elif isinstance(set, Set):
+            df = set.data
+        elif isinstance(set, range):
+            raise ValueError(
+                "Cannot convert a range to a set without a dimension name. Try Set(dim_name=range(...))"
+            )
+        else:
+            raise ValueError(f"Cannot convert type {type(set)} to a polars DataFrame")
+
+        if "index" in df.columns:
+            raise ValueError(
+                "Please specify a custom dimension name rather than using 'index' to avoid confusion."
+            )
+
+        for reserved_key in RESERVED_COL_KEYS:
+            if reserved_key in df.columns:
+                raise ValueError(
+                    f"Cannot use reserved column names {reserved_key} as dimensions."
+                )
+
+        return df
+
+
+class Expression(SupportsMath):
+    """Represents a linear or quadratic mathematical expression.
+
+    Examples:
+        >>> import pandas as pd
+        >>> df = pd.DataFrame(
+        ...     {
+        ...         "item": [1, 1, 1, 2, 2],
+        ...         "time": ["mon", "tue", "wed", "mon", "tue"],
+        ...         "cost": [1, 2, 3, 4, 5],
+        ...     }
+        ... ).set_index(["item", "time"])
+        >>> m = pf.Model()
+        >>> m.Time = pf.Variable(df.index)
+        >>> m.Size = pf.Variable(df.index)
+        >>> expr = df["cost"] * m.Time + df["cost"] * m.Size
+        >>> expr
+        <Expression height=5 terms=10 type=linear>
+        ┌──────┬──────┬──────────────────────────────┐
+        │ item ┆ time ┆ expression                   │
+        │ (2)  ┆ (3)  ┆                              │
+        ╞══════╪══════╪══════════════════════════════╡
+        │ 1    ┆ mon  ┆ Time[1,mon] + Size[1,mon]    │
+        │ 1    ┆ tue  ┆ 2 Time[1,tue] +2 Size[1,tue] │
+        │ 1    ┆ wed  ┆ 3 Time[1,wed] +3 Size[1,wed] │
+        │ 2    ┆ mon  ┆ 4 Time[2,mon] +4 Size[2,mon] │
+        │ 2    ┆ tue  ┆ 5 Time[2,tue] +5 Size[2,tue] │
+        └──────┴──────┴──────────────────────────────┘
+    """
+
+    def __init__(self, data: pl.DataFrame, name: str | None = None):
+        # Sanity checks, VAR_KEY and COEF_KEY must be present
+        assert VAR_KEY in data.columns, "Missing variable column."
+        assert COEF_KEY in data.columns, "Missing coefficient column."
+
+        # Sanity check no duplicates indices
+        if Config.enable_is_duplicated_expression_safety_check:
+            duplicated_mask = data.drop(COEF_KEY).is_duplicated()
+            # In theory this should never happen unless there's a bug in the library
+            if duplicated_mask.any():
+                duplicated_data = data.filter(duplicated_mask)
+                raise ValueError(
+                    f"Cannot create an expression with duplicate indices:\n{duplicated_data}."
+                )
+
+        data = _simplify_expr_df(data)
+
+        if name is None:
+            warnings.warn(
+                "Expression should be given a name to support troubleshooting.",
+                UserWarning,
+            )
+
+            super().__init__(data)
+        else:
+            super().__init__(data, name=name)
+
+    @classmethod
+    def constant(cls, constant: int | float) -> Expression:
+        """Creates a new expression equal to the given constant.
+
+        Examples:
+            >>> pf.Expression.constant(5)
+            <Expression terms=1 type=constant>
+            5
+        """
+        return cls(
+            pl.DataFrame(
+                {
+                    COEF_KEY: [constant],
+                    VAR_KEY: [CONST_TERM],
+                },
+                schema={COEF_KEY: pl.Float64, VAR_KEY: KEY_TYPE},
+            ),
+            name=str(constant),
+        )
+
+    @return_new
+    def sum(self, *over: str):
+        """Sums an expression over specified dimensions.
+
+        If no dimensions are specified, the sum is taken over all of the expression's dimensions.
+
+        Examples:
+            >>> expr = pl.DataFrame(
+            ...     {
+            ...         "time": ["mon", "tue", "wed", "mon", "tue"],
+            ...         "place": [
+            ...             "Toronto",
+            ...             "Toronto",
+            ...             "Toronto",
+            ...             "Vancouver",
+            ...             "Vancouver",
+            ...         ],
+            ...         "tiktok_posts": [1e6, 3e6, 2e6, 1e6, 2e6],
+            ...     }
+            ... ).to_expr()
+            >>> expr
+            <Expression height=5 terms=5 type=constant>
+            ┌──────┬───────────┬────────────┐
+            │ time ┆ place     ┆ expression │
+            │ (3)  ┆ (2)       ┆            │
+            ╞══════╪═══════════╪════════════╡
+            │ mon  ┆ Toronto   ┆ 1000000    │
+            │ tue  ┆ Toronto   ┆ 3000000    │
+            │ wed  ┆ Toronto   ┆ 2000000    │
+            │ mon  ┆ Vancouver ┆ 1000000    │
+            │ tue  ┆ Vancouver ┆ 2000000    │
+            └──────┴───────────┴────────────┘
+            >>> expr.sum("time")
+            <Expression height=2 terms=2 type=constant>
+            ┌───────────┬────────────┐
+            │ place     ┆ expression │
+            │ (2)       ┆            │
+            ╞═══════════╪════════════╡
+            │ Toronto   ┆ 6000000    │
+            │ Vancouver ┆ 3000000    │
+            └───────────┴────────────┘
+            >>> expr.sum()
+            <Expression terms=1 type=constant>
+            9000000
+
+            If the given dimensions don't exist, an error will be raised:
+
+            >>> expr.sum("city")
+            Traceback (most recent call last):
+            ...
+            AssertionError: Cannot sum over ['city'] as it is not in ['time', 'place']
+
+        See Also:
+            [pyoframe.Expression.sum_by][] for summing over all dimensions _except_ those that are specified.
+        """
+        dims = self.dimensions
+        if dims is None:
+            raise ValueError("Cannot sum a dimensionless expression.")
+        if not over:
+            over = tuple(dims)
+        assert set(over) <= set(dims), (
+            f"Cannot sum over {list(over)} as it is not in {dims}"
+        )
+        remaining_dims = [dim for dim in dims if dim not in over]
+
+        return (
+            self.data.drop(over)
+            .group_by(
+                remaining_dims + self._variable_columns,
+                maintain_order=Config.maintain_order,
+            )
+            .sum()
+        )
+
+    def sum_by(self, *by: str):
+        """Like [`Expression.sum`][pyoframe.Expression.sum], but the sum is taken over all dimensions *except* those specified in `by` (just like a `group_by().sum()` operation).
+
+        Examples:
+            >>> expr = pl.DataFrame(
+            ...     {
+            ...         "time": ["mon", "tue", "wed", "mon", "tue"],
+            ...         "place": [
+            ...             "Toronto",
+            ...             "Toronto",
+            ...             "Toronto",
+            ...             "Vancouver",
+            ...             "Vancouver",
+            ...         ],
+            ...         "tiktok_posts": [1e6, 3e6, 2e6, 1e6, 2e6],
+            ...     }
+            ... ).to_expr()
+            >>> expr
+            <Expression height=5 terms=5 type=constant>
+            ┌──────┬───────────┬────────────┐
+            │ time ┆ place     ┆ expression │
+            │ (3)  ┆ (2)       ┆            │
+            ╞══════╪═══════════╪════════════╡
+            │ mon  ┆ Toronto   ┆ 1000000    │
+            │ tue  ┆ Toronto   ┆ 3000000    │
+            │ wed  ┆ Toronto   ┆ 2000000    │
+            │ mon  ┆ Vancouver ┆ 1000000    │
+            │ tue  ┆ Vancouver ┆ 2000000    │
+            └──────┴───────────┴────────────┘
+
+            >>> expr.sum_by("place")
+            <Expression height=2 terms=2 type=constant>
+            ┌───────────┬────────────┐
+            │ place     ┆ expression │
+            │ (2)       ┆            │
+            ╞═══════════╪════════════╡
+            │ Toronto   ┆ 6000000    │
+            │ Vancouver ┆ 3000000    │
+            └───────────┴────────────┘
+
+            If the specified dimensions don't exist, an error will be raised:
+
+            >>> expr.sum_by("city")
+            Traceback (most recent call last):
+            ...
+            ValueError: Cannot sum by ['city'] because it is not a valid dimension. The expression's dimensions are: ['time', 'place'].
+
+            >>> total_sum = expr.sum()
+            >>> total_sum.sum_by("time")
+            Traceback (most recent call last):
+            ...
+            ValueError: Cannot sum a dimensionless expression.
+
+        See Also:
+            [pyoframe.Expression.sum][] for summing over specified dimensions.
+        """
+        if not by:
+            raise ValueError("sum_by requires at least 1 argument.")
+        dims = self.dimensions
+        if dims is None:
+            raise ValueError("Cannot sum a dimensionless expression.")
+        if not set(by) <= set(dims):
+            raise ValueError(
+                f"Cannot sum by {list(set(by) - set(dims))} because it is not a valid dimension. The expression's dimensions are: {list(dims)}."
+            )
+        remaining_dims = [dim for dim in dims if dim not in by]
+        return self.sum(*remaining_dims)
+
+    @property
+    def _variable_columns(self) -> list[str]:
+        if self.is_quadratic:
+            return [VAR_KEY, QUAD_VAR_KEY]
+        else:
+            return [VAR_KEY]
+
+    def map(self, mapping_set: SetTypes, drop_shared_dims: bool = True) -> Expression:
+        """Replaces the dimensions that are shared with mapping_set with the other dimensions found in mapping_set.
+
+        This is particularly useful to go from one type of dimensions to another. For example, to convert data that
+        is indexed by city to data indexed by country (see example).
+
+        Parameters:
+            mapping_set:
+                The set to map the expression to. This can be a DataFrame, Index, or another Set.
+            drop_shared_dims:
+                If `True`, the dimensions shared between the expression and the mapping set are dropped from the resulting expression and
+                    repeated rows are summed.
+                If `False`, the shared dimensions are kept in the resulting expression.
+
+        Returns:
+            A new Expression containing the result of the mapping operation.
+
+        Examples:
+            >>> import polars as pl
+            >>> pop_data = pl.DataFrame(
+            ...     {
+            ...         "city": ["Toronto", "Vancouver", "Boston"],
+            ...         "year": [2024, 2024, 2024],
+            ...         "population": [10, 2, 8],
+            ...     }
+            ... ).to_expr()
+            >>> cities_and_countries = pl.DataFrame(
+            ...     {
+            ...         "city": ["Toronto", "Vancouver", "Boston"],
+            ...         "country": ["Canada", "Canada", "USA"],
+            ...     }
+            ... )
+            >>> pop_data.map(cities_and_countries)
+            <Expression height=2 terms=2 type=constant>
+            ┌──────┬─────────┬────────────┐
+            │ year ┆ country ┆ expression │
+            │ (1)  ┆ (2)     ┆            │
+            ╞══════╪═════════╪════════════╡
+            │ 2024 ┆ Canada  ┆ 12         │
+            │ 2024 ┆ USA     ┆ 8          │
+            └──────┴─────────┴────────────┘
+
+            >>> pop_data.map(cities_and_countries, drop_shared_dims=False)
+            <Expression height=3 terms=3 type=constant>
+            ┌───────────┬──────┬─────────┬────────────┐
+            │ city      ┆ year ┆ country ┆ expression │
+            │ (3)       ┆ (1)  ┆ (2)     ┆            │
+            ╞═══════════╪══════╪═════════╪════════════╡
+            │ Toronto   ┆ 2024 ┆ Canada  ┆ 10         │
+            │ Vancouver ┆ 2024 ┆ Canada  ┆ 2          │
+            │ Boston    ┆ 2024 ┆ USA     ┆ 8          │
+            └───────────┴──────┴─────────┴────────────┘
+        """
+        mapping_set = Set(mapping_set)
+
+        dims = self.dimensions
+        if dims is None:
+            raise ValueError("Cannot use .map() on an expression with no dimensions.")
+
+        mapping_dims = mapping_set.dimensions
+        if mapping_dims is None:
+            raise ValueError(
+                "Cannot use .map() with a mapping set containing no dimensions."
+            )
+
+        shared_dims = [dim for dim in dims if dim in mapping_dims]
+        if not shared_dims:
+            raise ValueError(
+                f"Cannot apply .map() as there are no shared dimensions between the expression (dims={self.dimensions}) and the mapping set (dims={mapping_set.dimensions})."
+            )
+
+        mapped_expression = self * mapping_set
+
+        if drop_shared_dims:
+            mapped_expression = mapped_expression.sum(*shared_dims)
+
+        mapped_expression.name = f"{self.name}.map(…)"
+
+        return mapped_expression
+
+    @return_new
+    def rolling_sum(self, over: str, window_size: int):
+        """Calculates the rolling sum of the Expression over a specified window size for a given dimension.
+
+        This method applies a rolling sum operation over the dimension specified by `over`,
+        using a window defined by `window_size`.
+
+
+        Parameters:
+            over:
+                The name of the dimension (column) over which the rolling sum is calculated.
+                This dimension must exist within the Expression's dimensions.
+            window_size:
+                The size of the moving window in terms of number of records.
+                The rolling sum is calculated over this many consecutive elements.
+
+        Returns:
+            A new Expression instance containing the result of the rolling sum operation.
+                This new Expression retains all dimensions (columns) of the original data,
+                with the rolling sum applied over the specified dimension.
+
+        Examples:
+            >>> import polars as pl
+            >>> cost = pl.DataFrame(
+            ...     {
+            ...         "item": [1, 1, 1, 2, 2],
+            ...         "time": [1, 2, 3, 1, 2],
+            ...         "cost": [1, 2, 3, 4, 5],
+            ...     }
+            ... )
+            >>> m = pf.Model()
+            >>> m.quantity = pf.Variable(cost[["item", "time"]])
+            >>> (m.quantity * cost).rolling_sum(over="time", window_size=2)
+            <Expression height=5 terms=8 type=linear>
+            ┌──────┬──────┬──────────────────────────────────┐
+            │ item ┆ time ┆ expression                       │
+            │ (2)  ┆ (3)  ┆                                  │
+            ╞══════╪══════╪══════════════════════════════════╡
+            │ 1    ┆ 1    ┆ quantity[1,1]                    │
+            │ 1    ┆ 2    ┆ quantity[1,1] +2 quantity[1,2]   │
+            │ 1    ┆ 3    ┆ 2 quantity[1,2] +3 quantity[1,3] │
+            │ 2    ┆ 1    ┆ 4 quantity[2,1]                  │
+            │ 2    ┆ 2    ┆ 4 quantity[2,1] +5 quantity[2,2] │
+            └──────┴──────┴──────────────────────────────────┘
+        """
+        dims = self.dimensions
+        if dims is None:
+            raise ValueError(
+                "Cannot use rolling_sum() with an expression with no dimensions."
+            )
+        assert over in dims, f"Cannot sum over {over} as it is not in {dims}"
+        remaining_dims = [dim for dim in dims if dim not in over]
+
+        return pl.concat(
+            [
+                df.with_columns(pl.col(over).max())
+                for _, df in self.data.rolling(
+                    index_column=over,
+                    period=f"{window_size}i",
+                    group_by=remaining_dims,
+                )
+            ]
+        )
+
+    @return_new
+    def within(self, set: SetTypes):
+        """Filters this expression to only include the dimensions within the provided set.
+
+        Examples:
+            >>> import pandas as pd
+            >>> general_expr = pd.DataFrame(
+            ...     {"dim1": [1, 2, 3], "value": [1, 2, 3]}
+            ... ).to_expr()
+            >>> filter_expr = pd.DataFrame({"dim1": [1, 3], "value": [5, 6]}).to_expr()
+            >>> general_expr.within(filter_expr).data
+            shape: (2, 3)
+            ┌──────┬─────────┬───────────────┐
+            │ dim1 ┆ __coeff ┆ __variable_id │
+            │ ---  ┆ ---     ┆ ---           │
+            │ i64  ┆ f64     ┆ u32           │
+            ╞══════╪═════════╪═══════════════╡
+            │ 1    ┆ 1.0     ┆ 0             │
+            │ 3    ┆ 3.0     ┆ 0             │
+            └──────┴─────────┴───────────────┘
+        """
+        df: pl.DataFrame = Set(set).data
+        set_dims = _get_dimensions(df)
+        assert set_dims is not None, (
+            "Cannot use .within() with a set with no dimensions."
+        )
+        dims = self.dimensions
+        assert dims is not None, (
+            "Cannot use .within() with an expression with no dimensions."
+        )
+        dims_in_common = [dim for dim in dims if dim in set_dims]
+        by_dims = df.select(dims_in_common).unique(maintain_order=Config.maintain_order)
+        return self.data.join(
+            by_dims,
+            on=dims_in_common,
+            maintain_order="left" if Config.maintain_order else None,
+        )
+
+    @property
+    def is_quadratic(self) -> bool:
+        """Returns `True` if the expression is quadratic, False otherwise.
+
+        Computes in O(1) since expressions are quadratic if and
+        only if self.data contain the QUAD_VAR_KEY column.
+
+        Examples:
+            >>> import pandas as pd
+            >>> m = pf.Model()
+            >>> m.v = Variable()
+            >>> expr = pd.DataFrame({"dim1": [1, 2, 3], "value": [1, 2, 3]}) * m.v
+            >>> expr *= m.v
+            >>> expr.is_quadratic
+            True
+        """
+        return QUAD_VAR_KEY in self.data.columns
+
+    @overload
+    def degree(self, return_str: Literal[False] = False) -> int: ...
+
+    @overload
+    def degree(self, return_str: Literal[True] = True) -> str: ...
+
+    def degree(self, return_str: bool = False) -> int | str:
+        """Returns the degree of the expression (0=constant, 1=linear, 2=quadratic).
+
+        Parameters:
+            return_str: If `True`, returns the degree as a string (`"constant"`, `"linear"`, or `"quadratic"`).
+                If `False`, returns the degree as an integer (0, 1, or 2).
+
+        Examples:
+            >>> import pandas as pd
+            >>> m = pf.Model()
+            >>> m.v1 = pf.Variable()
+            >>> m.v2 = pf.Variable()
+            >>> expr = pd.DataFrame({"dim1": [1, 2, 3], "value": [1, 2, 3]}).to_expr()
+            >>> expr.degree()
+            0
+            >>> expr *= m.v1
+            >>> expr.degree()
+            1
+            >>> expr += (m.v2**2).over("dim1")
+            >>> expr.degree()
+            2
+            >>> expr.degree(return_str=True)
+            'quadratic'
+        """
+        if self.is_quadratic:
+            return "quadratic" if return_str else 2
+        # TODO improve performance of .evaluate() by ensuring early exit if linear
+        elif (self.data.get_column(VAR_KEY) != CONST_TERM).any():
+            return "linear" if return_str else 1
+        else:
+            return "constant" if return_str else 0
+
+    def __add__(self, other):
+        """Adds another expression or a constant to this expression.
+
+        Examples:
+            >>> import pandas as pd
+            >>> m = pf.Model()
+            >>> add = pd.DataFrame({"dim1": [1, 2, 3], "add": [10, 20, 30]}).to_expr()
+            >>> m.v = Variable(add)
+            >>> m.v + add
+            <Expression height=3 terms=6 type=linear>
+            ┌──────┬────────────┐
+            │ dim1 ┆ expression │
+            │ (3)  ┆            │
+            ╞══════╪════════════╡
+            │ 1    ┆ v[1] +10   │
+            │ 2    ┆ v[2] +20   │
+            │ 3    ┆ v[3] +30   │
+            └──────┴────────────┘
+
+            >>> m.v + add + 2
+            <Expression height=3 terms=6 type=linear>
+            ┌──────┬────────────┐
+            │ dim1 ┆ expression │
+            │ (3)  ┆            │
+            ╞══════╪════════════╡
+            │ 1    ┆ 12 + v[1]  │
+            │ 2    ┆ 22 + v[2]  │
+            │ 3    ┆ 32 + v[3]  │
+            └──────┴────────────┘
+
+            >>> m.v + pd.DataFrame({"dim1": [1, 2], "add": [10, 20]})
+            Traceback (most recent call last):
+            ...
+            pyoframe._constants.PyoframeError: Cannot add the two expressions below because of unmatched values.
+            Expression 1:   v
+            Expression 2:   add
+            Unmatched values:
+            shape: (1, 2)
+            ┌──────┬────────────┐
+            │ dim1 ┆ dim1_right │
+            │ ---  ┆ ---        │
+            │ i64  ┆ i64        │
+            ╞══════╪════════════╡
+            │ 3    ┆ null       │
+            └──────┴────────────┘
+            If this is intentional, use .drop_unmatched() or .keep_unmatched().
+            >>> m.v2 = Variable()
+            >>> 5 + 2 * m.v2
+            <Expression terms=2 type=linear>
+            2 v2 +5
+        """
+        if isinstance(other, (int, float)):
+            return self._add_const(other)
+        other = other.to_expr()
+        self._learn_from_other(other)
+        return add(self, other)
+
+    def __mul__(self: Expression, other: int | float | SupportsToExpr) -> Expression:
+        if isinstance(other, (int, float)):
+            if other == 1:
+                return self
+            return self._new(
+                self.data.with_columns(pl.col(COEF_KEY) * other),
+                name=f"({other} * {self.name})",
+            )
+
+        other = other.to_expr()
+        self._learn_from_other(other)
+        return multiply(self, other)
+
+    def to_expr(self) -> Expression:
+        """Returns the expression itself."""
+        return self
+
+    def _learn_from_other(self, other: Expression):
+        if self._model is None and other._model is not None:
+            self._model = other._model
+
+    def _new(self, data: pl.DataFrame, name: str) -> Expression:
+        e = Expression(data, name)
+        e._model = self._model
+        return e
+
+    def _add_const(self, const: int | float) -> Expression:
+        """Adds a constant to the expression.
+
+        Examples:
+            >>> m = pf.Model()
+            >>> m.x1 = Variable()
+            >>> m.x2 = Variable()
+            >>> m.x1 + 5
+            <Expression terms=2 type=linear>
+            x1 +5
+            >>> m.x1**2 + 5
+            <Expression terms=2 type=quadratic>
+            x1 * x1 +5
+            >>> m.x1**2 + m.x2 + 5
+            <Expression terms=3 type=quadratic>
+            x1 * x1 + x2 +5
+
+            It also works with dimensions
+
+            >>> m = pf.Model()
+            >>> m.v = Variable({"dim1": [1, 2, 3]})
+            >>> m.v * m.v + 5
+            <Expression height=3 terms=6 type=quadratic>
+            ┌──────┬─────────────────┐
+            │ dim1 ┆ expression      │
+            │ (3)  ┆                 │
+            ╞══════╪═════════════════╡
+            │ 1    ┆ 5 + v[1] * v[1] │
+            │ 2    ┆ 5 + v[2] * v[2] │
+            │ 3    ┆ 5 + v[3] * v[3] │
+            └──────┴─────────────────┘
+        """
+        if const == 0:
+            return self
+        dim = self.dimensions
+        data = self.data
+        # Fill in missing constant terms
+        if not dim:
+            if CONST_TERM not in data[VAR_KEY]:
+                const_df = pl.DataFrame(
+                    {COEF_KEY: [0.0], VAR_KEY: [CONST_TERM]},
+                    schema={COEF_KEY: pl.Float64, VAR_KEY: KEY_TYPE},
+                )
+                if self.is_quadratic:
+                    const_df = const_df.with_columns(
+                        pl.lit(CONST_TERM).alias(QUAD_VAR_KEY).cast(KEY_TYPE)
+                    )
+                data = pl.concat(
+                    [data, const_df],
+                    how="vertical_relaxed",
+                )
+        else:
+            keys = (
+                data.select(dim)
+                .unique(maintain_order=Config.maintain_order)
+                .with_columns(pl.lit(CONST_TERM).alias(VAR_KEY).cast(KEY_TYPE))
+            )
+            if self.is_quadratic:
+                keys = keys.with_columns(
+                    pl.lit(CONST_TERM).alias(QUAD_VAR_KEY).cast(KEY_TYPE)
+                )
+            data = data.join(
+                keys,
+                on=dim + self._variable_columns,
+                how="full",
+                coalesce=True,
+                # We use right_left not left_right to bring the constants near the front for better readability
+                maintain_order="right_left" if Config.maintain_order else None,
+            ).with_columns(pl.col(COEF_KEY).fill_null(0.0))
+
+        data = data.with_columns(
+            pl.when(pl.col(VAR_KEY) == CONST_TERM)
+            .then(pl.col(COEF_KEY) + const)
+            .otherwise(pl.col(COEF_KEY))
+        )
+
+        name = f"({self.name} + {const})" if const >= 0 else f"({self.name} - {-const})"
+        return self._new(data, name=name)
+
+    @property
+    def constant_terms(self) -> pl.DataFrame:
+        """Returns all the constant terms in the expression."""
+        dims = self.dimensions
+        constant_terms = self.data.filter(pl.col(VAR_KEY) == CONST_TERM).drop(VAR_KEY)
+        if self.is_quadratic:
+            constant_terms = constant_terms.drop(QUAD_VAR_KEY)
+        if dims is not None:
+            dims_df = self.data.select(dims).unique(
+                maintain_order=Config.maintain_order
+            )
+            df = constant_terms.join(
+                dims_df,
+                on=dims,
+                how="full",
+                coalesce=True,
+                maintain_order="left_right" if Config.maintain_order else None,
+            )
+            return df.with_columns(pl.col(COEF_KEY).fill_null(0.0))
+        else:
+            if len(constant_terms) == 0:
+                return pl.DataFrame(
+                    {COEF_KEY: [0.0], VAR_KEY: [CONST_TERM]},
+                    schema={COEF_KEY: pl.Float64, VAR_KEY: KEY_TYPE},
+                )
+            return constant_terms
+
+    @property
+    def variable_terms(self) -> pl.DataFrame:
+        """Returns all the non-constant terms in the expression."""
+        return self.data.filter(pl.col(VAR_KEY) != CONST_TERM)
+
+    @unwrap_single_values
+    def evaluate(self) -> pl.DataFrame:
+        """Computes the value of the expression using the variables' solutions.
+
+        Returns:
+            A Polars `DataFrame` for dimensioned expressions a `float` for dimensionless expressions.
+
+        Examples:
+            >>> m = pf.Model()
+            >>> m.X = pf.Variable({"dim1": [1, 2, 3]}, lb=10, ub=10)
+            >>> m.expr = 2 * m.X * m.X + 1
+
+            >>> m.expr.evaluate()
+            Traceback (most recent call last):
+            ...
+            ValueError: Cannot evaluate the expression 'expr' before calling model.optimize().
+
+            >>> m.constant_expression = m.expr - 2 * m.X * m.X
+            >>> m.constant_expression.evaluate()
+            shape: (3, 2)
+            ┌──────┬──────────┐
+            │ dim1 ┆ solution │
+            │ ---  ┆ ---      │
+            │ i64  ┆ f64      │
+            ╞══════╪══════════╡
+            │ 1    ┆ 1.0      │
+            │ 2    ┆ 1.0      │
+            │ 3    ┆ 1.0      │
+            └──────┴──────────┘
+
+
+            >>> m.optimize()
+            >>> m.expr.evaluate()
+            shape: (3, 2)
+            ┌──────┬──────────┐
+            │ dim1 ┆ solution │
+            │ ---  ┆ ---      │
+            │ i64  ┆ f64      │
+            ╞══════╪══════════╡
+            │ 1    ┆ 201.0    │
+            │ 2    ┆ 201.0    │
+            │ 3    ┆ 201.0    │
+            └──────┴──────────┘
+
+            >>> m.expr.sum().evaluate()
+            603.0
+
+        """
+        assert self._model is not None, (
+            "Expression must be added to the model to use .value"
+        )
+
+        df = self.data.rename({COEF_KEY: SOLUTION_KEY})
+        sm = self._model.poi
+        attr = poi.VariableAttribute.Value
+
+        if self.degree() == 0:
+            df = df.drop(self._variable_columns)
+        elif (
+            self._model.attr.TerminationStatus
+            == poi.TerminationStatusCode.OPTIMIZE_NOT_CALLED
+        ):
+            raise ValueError(
+                f"Cannot evaluate the expression '{self.name}' before calling model.optimize()."
+            )
+        else:
+            for var_col in self._variable_columns:
+                values = [
+                    sm.get_variable_attribute(poi.VariableIndex(v_id), attr)
+                    for v_id in df.get_column(var_col).to_list()
+                ]
+
+                df = df.drop(var_col).with_columns(
+                    pl.col(SOLUTION_KEY) * pl.Series(values, dtype=pl.Float64)
+                )
+
+        dims = self.dimensions
+        if dims is not None:
+            df = df.group_by(dims, maintain_order=Config.maintain_order)
+        return df.sum()
+
+    def _to_poi(self) -> poi.ScalarAffineFunction | poi.ScalarQuadraticFunction:
+        assert self.dimensions is None, (
+            "._to_poi() only works for non-dimensioned expressions."
+        )
+
+        if self.is_quadratic:
+            return poi.ScalarQuadraticFunction(
+                coefficients=self.data.get_column(COEF_KEY).to_numpy(),
+                var1s=self.data.get_column(VAR_KEY).to_numpy(),
+                var2s=self.data.get_column(QUAD_VAR_KEY).to_numpy(),
+            )
+        else:
+            return poi.ScalarAffineFunction(
+                coefficients=self.data.get_column(COEF_KEY).to_numpy(),
+                variables=self.data.get_column(VAR_KEY).to_numpy(),
+            )
+
+    @overload
+    def to_str(
+        self,
+        str_col_name: str = "expression",
+        include_const_term: bool = True,
+        return_df: Literal[False] = False,
+    ) -> str: ...
+
+    @overload
+    def to_str(
+        self,
+        str_col_name: str = "expression",
+        include_const_term: bool = True,
+        return_df: Literal[True] = True,
+    ) -> pl.DataFrame: ...
+
+    def to_str(
+        self,
+        str_col_name: str = "expression",
+        include_const_term: bool = True,
+        return_df: bool = False,
+    ) -> str | pl.DataFrame:
+        """Converts the expression to a human-readable string, or several arranged in a table.
+
+        Long expressions are truncated according to [`Config.print_max_terms`][pyoframe._Config.print_max_terms] and [`Config.print_polars_config`][pyoframe._Config.print_polars_config].
+
+        `str(pyoframe.Expression)` is equivalent to `pyoframe.Expression.to_str()`.
+
+        Parameters:
+            str_col_name:
+                The name of the column containing the string representation of the expression (dimensioned expressions only).
+            include_const_term:
+                If `False`, constant terms are omitted from the string representation.
+            return_df:
+                If `True`, returns a DataFrame containing the human-readable strings instead of the DataFrame's string representation.
+
+        Examples:
+            >>> import polars as pl
+            >>> m = pf.Model()
+            >>> x = pf.Set(x=range(1000))
+            >>> y = pf.Set(y=range(1000))
+            >>> m.V = pf.Variable(x, y)
+            >>> expr = 2 * m.V * m.V + 3
+            >>> print(expr.to_str())
+            ┌────────┬────────┬──────────────────────────────┐
+            │ x      ┆ y      ┆ expression                   │
+            │ (1000) ┆ (1000) ┆                              │
+            ╞════════╪════════╪══════════════════════════════╡
+            │ 0      ┆ 0      ┆ 3 +2 V[0,0] * V[0,0]         │
+            │ 0      ┆ 1      ┆ 3 +2 V[0,1] * V[0,1]         │
+            │ 0      ┆ 2      ┆ 3 +2 V[0,2] * V[0,2]         │
+            │ 0      ┆ 3      ┆ 3 +2 V[0,3] * V[0,3]         │
+            │ 0      ┆ 4      ┆ 3 +2 V[0,4] * V[0,4]         │
+            │ …      ┆ …      ┆ …                            │
+            │ 999    ┆ 995    ┆ 3 +2 V[999,995] * V[999,995] │
+            │ 999    ┆ 996    ┆ 3 +2 V[999,996] * V[999,996] │
+            │ 999    ┆ 997    ┆ 3 +2 V[999,997] * V[999,997] │
+            │ 999    ┆ 998    ┆ 3 +2 V[999,998] * V[999,998] │
+            │ 999    ┆ 999    ┆ 3 +2 V[999,999] * V[999,999] │
+            └────────┴────────┴──────────────────────────────┘
+            >>> expr = expr.sum("y")
+            >>> print(expr.to_str())
+            ┌────────┬─────────────────────────────────────────────────────────────────────────────────────────┐
+            │ x      ┆ expression                                                                              │
+            │ (1000) ┆                                                                                         │
+            ╞════════╪═════════════════════════════════════════════════════════════════════════════════════════╡
+            │ 0      ┆ 3000 +2 V[0,0] * V[0,0] +2 V[0,1] * V[0,1] +2 V[0,2] * V[0,2] +2 V[0,3] * V[0,3] …      │
+            │ 1      ┆ 3000 +2 V[1,0] * V[1,0] +2 V[1,1] * V[1,1] +2 V[1,2] * V[1,2] +2 V[1,3] * V[1,3] …      │
+            │ 2      ┆ 3000 +2 V[2,0] * V[2,0] +2 V[2,1] * V[2,1] +2 V[2,2] * V[2,2] +2 V[2,3] * V[2,3] …      │
+            │ 3      ┆ 3000 +2 V[3,0] * V[3,0] +2 V[3,1] * V[3,1] +2 V[3,2] * V[3,2] +2 V[3,3] * V[3,3] …      │
+            │ 4      ┆ 3000 +2 V[4,0] * V[4,0] +2 V[4,1] * V[4,1] +2 V[4,2] * V[4,2] +2 V[4,3] * V[4,3] …      │
+            │ …      ┆ …                                                                                       │
+            │ 995    ┆ 3000 +2 V[995,0] * V[995,0] +2 V[995,1] * V[995,1] +2 V[995,2] * V[995,2] +2 V[995,3] * │
+            │        ┆ V[995,3] …                                                                              │
+            │ 996    ┆ 3000 +2 V[996,0] * V[996,0] +2 V[996,1] * V[996,1] +2 V[996,2] * V[996,2] +2 V[996,3] * │
+            │        ┆ V[996,3] …                                                                              │
+            │ 997    ┆ 3000 +2 V[997,0] * V[997,0] +2 V[997,1] * V[997,1] +2 V[997,2] * V[997,2] +2 V[997,3] * │
+            │        ┆ V[997,3] …                                                                              │
+            │ 998    ┆ 3000 +2 V[998,0] * V[998,0] +2 V[998,1] * V[998,1] +2 V[998,2] * V[998,2] +2 V[998,3] * │
+            │        ┆ V[998,3] …                                                                              │
+            │ 999    ┆ 3000 +2 V[999,0] * V[999,0] +2 V[999,1] * V[999,1] +2 V[999,2] * V[999,2] +2 V[999,3] * │
+            │        ┆ V[999,3] …                                                                              │
+            └────────┴─────────────────────────────────────────────────────────────────────────────────────────┘
+            >>> expr = expr.sum("x")
+            >>> print(expr.to_str())
+            3000000 +2 V[0,0] * V[0,0] +2 V[0,1] * V[0,1] +2 V[0,2] * V[0,2] +2 V[0,3] * V[0,3] …
+
+        """
+        # TODO consider optimizing using LazyFrames since .head() could maybe be automatically pushed up the chain of operations.
+        data = self.data if include_const_term else self.variable_terms
+        data = cast_coef_to_string(data)
+
+        for var_col in self._variable_columns:
+            temp_var_column = f"{var_col}_temp"
+            if self._model is not None and self._model._var_map is not None:
+                data = self._model._var_map.apply(
+                    data, to_col=temp_var_column, id_col=var_col
+                )
+            else:
+                data = data.with_columns(
+                    pl.concat_str(pl.lit("x"), var_col).alias(temp_var_column)
+                )
+            data = data.with_columns(
+                pl.when(pl.col(var_col) == CONST_TERM)
+                .then(pl.lit(""))
+                .otherwise(temp_var_column)
+                .alias(var_col)
+            ).drop(temp_var_column)
+        if self.is_quadratic:
+            data = data.with_columns(
+                pl.when(pl.col(QUAD_VAR_KEY) == "")
+                .then(pl.col(VAR_KEY))
+                .otherwise(pl.concat_str(VAR_KEY, pl.lit(" * "), pl.col(QUAD_VAR_KEY)))
+                .alias(VAR_KEY)
+            ).drop(QUAD_VAR_KEY)
+
+        dimensions = self.dimensions
+
+        # Create a string for each term
+        data = data.with_columns(
+            pl.concat_str(
+                COEF_KEY,
+                pl.lit(" "),
+                VAR_KEY,
+            )
+            .str.strip_chars(characters="  ")
+            .alias(str_col_name)
+        ).drop(COEF_KEY, VAR_KEY)
+
+        if dimensions is not None:
+            data = data.group_by(dimensions, maintain_order=Config.maintain_order).agg(
+                pl.concat_str(
+                    pl.col(str_col_name)
+                    .head(Config.print_max_terms)
+                    .str.join(delimiter=" "),
+                    pl.when(pl.len() > Config.print_max_terms)
+                    .then(pl.lit(" …"))
+                    .otherwise(pl.lit("")),
+                )
+            )
+        else:
+            truncate = data.height > Config.print_max_terms
+            if truncate:
+                data = data.head(Config.print_max_terms)
+
+            data = data.select(pl.col(str_col_name).str.join(delimiter=" "))
+
+            if truncate:
+                data = data.with_columns(
+                    pl.concat_str(pl.col(str_col_name), pl.lit(" …"))
+                )
+
+        # Remove leading +
+        data = data.with_columns(pl.col(str_col_name).str.strip_chars(characters="  +"))
+
+        if not return_df:
+            if dimensions is None and not self._allowed_new_dims:
+                data = data.item()
+            else:
+                data = self._add_shape_to_columns(data)
+                data = self._add_allowed_new_dims_to_df(data)
+                with Config.print_polars_config:
+                    data = repr(data)
+
+        return data
+
+    def _str_header(self) -> str:
+        """Returns a string representation of the expression's header."""
+        return get_obj_repr(
+            self,
+            height=len(self) if self.dimensions else None,
+            terms=self.terms,
+            type=self.degree(return_str=True),
+        )
+
+    def __repr__(self) -> str:
+        return self._str_header() + "\n" + self.to_str()
+
+    def __str__(self) -> str:
+        return self.to_str()
+
+    @property
+    def terms(self) -> int:
+        """The number of terms across all subexpressions.
+
+        Expressions equal to zero count as one term.
+
+        Examples:
+            >>> import polars as pl
+            >>> m = pf.Model()
+            >>> m.v = pf.Variable({"t": [1, 2]})
+            >>> coef = pl.DataFrame({"t": [1, 2], "coef": [0, 1]})
+            >>> coef * (m.v + 4)
+            <Expression height=2 terms=3 type=linear>
+            ┌─────┬────────────┐
+            │ t   ┆ expression │
+            │ (2) ┆            │
+            ╞═════╪════════════╡
+            │ 1   ┆ 0          │
+            │ 2   ┆ 4 + v[2]   │
+            └─────┴────────────┘
+            >>> (coef * (m.v + 4)).terms
+            3
+        """
+        return len(self.data)
+
+
+@overload
+def sum(over: str | Sequence[str], expr: SupportsToExpr) -> Expression: ...
+
+
+@overload
+def sum(over: SupportsToExpr) -> Expression: ...
+
+
+def sum(
+    over: str | Sequence[str] | SupportsToExpr,
+    expr: SupportsToExpr | None = None,
+) -> Expression:  # pragma: no cover
+    """Deprecated: Use Expression.sum() or Variable.sum() instead.
+
+    Examples:
+        >>> x = pf.Set(x=range(100))
+        >>> pf.sum(x)
+        Traceback (most recent call last):
+          ...
+        DeprecationWarning: pf.sum() is deprecated. Use Expression.sum() or Variable.sum() instead.
+    """
+    warnings.warn(
+        "pf.sum() is deprecated. Use Expression.sum() or Variable.sum() instead.",
+        DeprecationWarning,
+    )
+
+    if expr is None:
+        assert isinstance(over, SupportsMath)
+        return over.to_expr().sum()
+    else:
+        assert isinstance(over, (str, Sequence))
+        if isinstance(over, str):
+            over = (over,)
+        return expr.to_expr().sum(*over)
+
+
+def sum_by(
+    by: str | Sequence[str], expr: SupportsToExpr
+) -> Expression:  # pragma: no cover
+    """Deprecated: Use Expression.sum() or Variable.sum() instead."""
+    warnings.warn(
+        "pf.sum_by() is deprecated. Use Expression.sum_by() or Variable.sum_by() instead.",
+        DeprecationWarning,
+    )
+
+    if isinstance(by, str):
+        by = [by]
+    return expr.to_expr().sum_by(*by)
+
+
+class Constraint(ModelElementWithId):
+    """An optimization constraint that can be added to a [Model][pyoframe.Model].
+
+    Tip: Implementation Note
+        Pyoframe simplifies constraints by moving all the constraint's mathematical terms to the left-hand side.
+        This way, the right-hand side is always zero, and constraints only need to manage one expression.
+
+    Warning: Use `<=`, `>=`, or `==` operators to create constraints
+        Constraints should be created using the `<=`, `>=`, or `==` operators, not by directly calling the `Constraint` constructor.
+
+    Parameters:
+        lhs:
+            The constraint's left-hand side expression.
+        sense:
+            The sense of the constraint.
+    """
+
+    def __init__(self, lhs: Expression, sense: ConstraintSense):
+        self.lhs: Expression = lhs
+        self._model = lhs._model
+        self.sense = sense
+        self._to_relax: FuncArgs | None = None
+        self._attr = Container(self._set_attribute, self._get_attribute)
+
+        dims = self.lhs.dimensions
+        data = (
+            pl.DataFrame()
+            if dims is None
+            else self.lhs.data.select(dims).unique(maintain_order=Config.maintain_order)
+        )
+
+        super().__init__(data)
+
+    @property
+    def attr(self) -> Container:
+        """Allows reading and writing constraint attributes similarly to [Model.attr][pyoframe.Model.attr]."""
+        return self._attr
+
+    def _set_attribute(self, name, value):
+        self._assert_has_ids()
+        col_name = name
+        try:
+            name = poi.ConstraintAttribute[name]
+            setter = self._model.poi.set_constraint_attribute
+        except KeyError:
+            setter = self._model.poi.set_constraint_raw_attribute
+
+        constr_type = (
+            poi.ConstraintType.Quadratic
+            if self.lhs.is_quadratic
+            else poi.ConstraintType.Linear
+        )
+
+        if self.dimensions is None:
+            for key in self.data.get_column(CONSTRAINT_KEY):
+                setter(poi.ConstraintIndex(constr_type, key), name, value)
+        else:
+            for key, value in (
+                self.data.join(
+                    value,
+                    on=self.dimensions,
+                    maintain_order="left" if Config.maintain_order else None,
+                )
+                .select(pl.col(CONSTRAINT_KEY), pl.col(col_name))
+                .iter_rows()
+            ):
+                setter(poi.ConstraintIndex(constr_type, key), name, value)
+
+    @unwrap_single_values
+    def _get_attribute(self, name):
+        self._assert_has_ids()
+        col_name = name
+        try:
+            name = poi.ConstraintAttribute[name]
+            getter = self._model.poi.get_constraint_attribute
+        except KeyError:
+            getter = self._model.poi.get_constraint_raw_attribute
+
+        constr_type = (
+            poi.ConstraintType.Quadratic
+            if self.lhs.is_quadratic
+            else poi.ConstraintType.Linear
+        )
+
+        ids = self.data.get_column(CONSTRAINT_KEY).to_list()
+        attr = [getter(poi.ConstraintIndex(constr_type, v_id), name) for v_id in ids]
+        data = self.data.with_columns(pl.Series(attr).alias(col_name))
+        return data.select(self._dimensions_unsafe + [col_name])
+
+    def _on_add_to_model(self, model: Model, name: str):
+        super()._on_add_to_model(model, name)
+        if self._to_relax is not None:
+            self.relax(*self._to_relax.args, **self._to_relax.kwargs)
+        self._assign_ids()
+
+    def _assign_ids(self):
+        """This function is the main bottleneck for pyoframe.
+
+        I've spent a lot of time optimizing it.
+        """
+        assert self._model is not None
+
+        is_quadratic = self.lhs.is_quadratic
+        use_var_names = self._model.solver_uses_variable_names
+        sense = self.sense._to_poi()
+        dims = self.dimensions
+        df = self.lhs.data
+        add_constraint = (
+            self._model.poi._add_quadratic_constraint
+            if is_quadratic
+            else self._model.poi._add_linear_constraint
+        )
+
+        # GRBaddconstr uses sprintf when no name or "" is given. sprintf is slow. As such, we specify "C" as the name.
+        # Specifying "" is the same as not specifying anything, see pyoptinterface:
+        # https://github.com/metab0t/PyOptInterface/blob/6d61f3738ad86379cff71fee77077d4ea919f2d5/lib/gurobi_model.cpp#L338
+        name = "C" if self._model.solver.block_auto_names else ""
+
+        if dims is None:
+            if self._model.solver_uses_variable_names:
+                name = self.name
+            create_expression = (
+                poi.ScalarQuadraticFunction
+                if is_quadratic
+                else poi.ScalarAffineFunction.from_numpy  # when called only once from_numpy is faster
+            )
+            df = self.data.with_columns(
+                pl.lit(
+                    add_constraint(
+                        create_expression(
+                            *(
+                                df.get_column(c).to_numpy()
+                                for c in ([COEF_KEY] + self.lhs._variable_columns)
+                            )
+                        ),
+                        sense,
+                        0,
+                        name,
+                    ).index
+                )
+                .alias(CONSTRAINT_KEY)
+                .cast(KEY_TYPE)
+            )
+        else:
+            create_expression = (
+                poi.ScalarQuadraticFunction
+                if is_quadratic
+                else poi.ScalarAffineFunction  # when called multiple times the default constructor is fastest
+            )
+            if Config.maintain_order:
+                # This adds a 5-10% overhead on _assign_ids but ensures the order
+                # is the same as the input data
+                df_unique = df.select(dims).unique(maintain_order=True)
+                df = (
+                    df.join(
+                        df_unique.with_row_index(),
+                        on=dims,
+                        maintain_order="left",
+                    )
+                    .sort("index", maintain_order=True)
+                    .drop("index")
+                )
+            else:
+                df = df.sort(dims, maintain_order=False)
+                # must maintain order otherwise results are wrong!
+                df_unique = df.select(dims).unique(maintain_order=True)
+            coefs = df.get_column(COEF_KEY).to_list()
+            vars = df.get_column(VAR_KEY).to_list()
+            if is_quadratic:
+                vars2 = df.get_column(QUAD_VAR_KEY).to_list()
+
+            split = (
+                df.lazy()
+                .with_row_index()
+                .filter(pl.struct(dims).is_first_distinct())
+                .select("index")
+                .collect()
+                .to_series()
+                .to_list()
+            ) + [df.height]
+            del df
+
+            # Note: list comprehension was slightly faster than using polars map_elements
+            # Note 2: not specifying the argument name (`expr=`) was also slightly faster.
+            # Note 3: we could have merged the if-else using an expansion operator (*) but that is slow.
+            # Note 4: using kwargs is slow and including the constant term for linear expressions is faster.
+            if use_var_names:
+                names = concat_dimensions(df_unique, prefix=self.name)[
+                    "concated_dim"
+                ].to_list()
+                if is_quadratic:
+                    ids = [
+                        add_constraint(
+                            create_expression(coefs[s0:s1], vars[s0:s1], vars2[s0:s1]),
+                            sense,
+                            0,
+                            names[i],
+                        ).index
+                        for i, (s0, s1) in enumerate(pairwise(split))
+                    ]
+                else:
+                    ids = [
+                        add_constraint(
+                            create_expression(coefs[s0:s1], vars[s0:s1], 0),
+                            sense,
+                            0,
+                            names[i],
+                        ).index
+                        for i, (s0, s1) in enumerate(pairwise(split))
+                    ]
+            else:
+                if is_quadratic:
+                    ids = [
+                        add_constraint(
+                            create_expression(coefs[s0:s1], vars[s0:s1], vars2[s0:s1]),
+                            sense,
+                            0,
+                            name,
+                        ).index
+                        for s0, s1 in pairwise(split)
+                    ]
+                else:
+                    ids = [
+                        add_constraint(
+                            create_expression(coefs[s0:s1], vars[s0:s1], 0),
+                            sense,
+                            0,
+                            name,
+                        ).index
+                        for s0, s1 in pairwise(split)
+                    ]
+            df = df_unique.with_columns(
+                pl.Series(ids, dtype=KEY_TYPE).alias(CONSTRAINT_KEY)
+            )
+
+        self._data = df
+
+    @property
+    def dual(self) -> pl.DataFrame | float:
+        """Returns the constraint's dual values.
+
+        Examples:
+            >>> m = pf.Model()
+            >>> m.x = pf.Variable()
+            >>> m.y = pf.Variable()
+            >>> m.maximize = m.x - m.y
+
+            Notice that for every unit increase in the right-hand side, the objective only improves by 0.5.
+            >>> m.constraint_x = 2 * m.x <= 10
+            >>> m.constraint_y = 2 * m.y >= 5
+            >>> m.optimize()
+
+            For every unit increase in the right-hand side of `constraint_x`, the objective improves by 0.5.
+            >>> m.constraint_x.dual
+            0.5
+
+            For every unit increase in the right-hand side of `constraint_y`, the objective worsens by 0.5.
+            >>> m.constraint_y.dual
+            -0.5
+        """
+        dual = self.attr.Dual
+        if isinstance(dual, pl.DataFrame):
+            dual = dual.rename({"Dual": DUAL_KEY})
+
+        # Weirdly, IPOPT returns dual values with the opposite sign, so we correct this bug.
+        # It also does this for maximization problems
+        # but since we flip the objective (because Ipopt doesn't support maximization), the double negatives cancel out.
+        assert self._model is not None
+        if self._model.solver.name == "ipopt" and self._model.sense == ObjSense.MIN:
+            if isinstance(dual, pl.DataFrame):
+                dual = dual.with_columns(-pl.col(DUAL_KEY))
+            else:
+                dual = -dual
+        return dual
+
+    @classmethod
+    def _get_id_column_name(cls):
+        return CONSTRAINT_KEY
+
+    def filter(self, *args, **kwargs) -> pl.DataFrame:
+        """Syntactic sugar on `Constraint.lhs.data.filter()`, to help debugging."""
+        return self.lhs.data.filter(*args, **kwargs)
+
+    def relax(
+        self, cost: SupportsToExpr, max: SupportsToExpr | None = None
+    ) -> Constraint:
+        """Allows the constraint to be violated at a `cost` and, optionally, up to a maximum.
+
+        Warning:
+            `.relax()` must be called before the constraint is assigned to the [Model][pyoframe.Model] (see examples below).
+
+        Parameters:
+            cost:
+                The cost of violating the constraint. Costs should be positive because Pyoframe will automatically
+                make them negative for maximization problems.
+            max:
+                The maximum value of the relaxation variable.
+
+        Returns:
+            The same constraint
+
+        Examples:
+            >>> m = pf.Model()
+            >>> m.hours_sleep = pf.Variable(lb=0)
+            >>> m.hours_day = pf.Variable(lb=0)
+            >>> m.hours_in_day = m.hours_sleep + m.hours_day == 24
+            >>> m.maximize = m.hours_day
+            >>> m.must_sleep = (m.hours_sleep >= 8).relax(cost=2, max=3)
+            >>> m.optimize()
+            >>> m.hours_day.solution
+            16.0
+            >>> m.maximize += 2 * m.hours_day
+            >>> m.optimize()
+            >>> m.hours_day.solution
+            19.0
+
+            `relax` can only be called after the sense of the model has been defined.
+
+            >>> m = pf.Model()
+            >>> m.hours_sleep = pf.Variable(lb=0)
+            >>> m.hours_day = pf.Variable(lb=0)
+            >>> m.hours_in_day = m.hours_sleep + m.hours_day == 24
+            >>> m.must_sleep = (m.hours_sleep >= 8).relax(cost=2, max=3)
+            Traceback (most recent call last):
+            ...
+            ValueError: Cannot relax a constraint before the objective sense has been set. Try setting the objective first or using Model(sense=...).
+
+            One way to solve this is by setting the sense directly on the model. See how this works fine:
+
+            >>> m = pf.Model(sense="max")
+            >>> m.hours_sleep = pf.Variable(lb=0)
+            >>> m.hours_day = pf.Variable(lb=0)
+            >>> m.hours_in_day = m.hours_sleep + m.hours_day == 24
+            >>> m.must_sleep = (m.hours_sleep >= 8).relax(cost=2, max=3)
+
+            And now an example with dimensions:
+
+            >>> homework_due_tomorrow = pl.DataFrame(
+            ...     {
+            ...         "project": ["A", "B", "C"],
+            ...         "cost_per_hour_underdelivered": [10, 20, 30],
+            ...         "hours_to_finish": [9, 9, 9],
+            ...         "max_underdelivered": [1, 9, 9],
+            ...     }
+            ... )
+            >>> m.hours_spent = pf.Variable(homework_due_tomorrow["project"], lb=0)
+            >>> m.must_finish_project = (
+            ...     m.hours_spent
+            ...     >= homework_due_tomorrow[["project", "hours_to_finish"]]
+            ... ).relax(
+            ...     homework_due_tomorrow[["project", "cost_per_hour_underdelivered"]],
+            ...     max=homework_due_tomorrow[["project", "max_underdelivered"]],
+            ... )
+            >>> m.only_one_day = m.hours_spent.sum("project") <= 24
+            >>> # Relaxing a constraint after it has already been assigned will give an error
+            >>> m.only_one_day.relax(1)
+            Traceback (most recent call last):
+            ...
+            ValueError: .relax() must be called before the Constraint is added to the model
+            >>> m.attr.Silent = True
+            >>> m.optimize()
+            >>> m.maximize.value
+            -50.0
+            >>> m.hours_spent.solution
+            shape: (3, 2)
+            ┌─────────┬──────────┐
+            │ project ┆ solution │
+            │ ---     ┆ ---      │
+            │ str     ┆ f64      │
+            ╞═════════╪══════════╡
+            │ A       ┆ 8.0      │
+            │ B       ┆ 7.0      │
+            │ C       ┆ 9.0      │
+            └─────────┴──────────┘
+        """
+        if self._has_ids:
+            raise ValueError(
+                ".relax() must be called before the Constraint is added to the model"
+            )
+
+        m = self._model
+        if m is None:
+            self._to_relax = FuncArgs(args=[cost, max])
+            return self
+
+        var_name = f"{self.name}_relaxation"
+        assert not hasattr(m, var_name), (
+            "Conflicting names, relaxation variable already exists on the model."
+        )
+        var = Variable(self, lb=0, ub=max)
+        setattr(m, var_name, var)
+
+        if self.sense == ConstraintSense.LE:
+            self.lhs -= var
+        elif self.sense == ConstraintSense.GE:
+            self.lhs += var
+        else:  # pragma: no cover
+            # TODO
+            raise NotImplementedError(
+                "Relaxation for equalities has not yet been implemented. Submit a pull request!"
+            )
+
+        penalty = var * cost
+        if self.dimensions:
+            penalty = penalty.sum()
+        if m.sense is None:
+            raise ValueError(
+                "Cannot relax a constraint before the objective sense has been set. Try setting the objective first or using Model(sense=...)."
+            )
+        elif m.sense == ObjSense.MAX:
+            penalty *= -1
+        if m.has_objective:
+            m.objective += penalty
+        else:
+            m.objective = penalty
+
+        return self
+
+    def estimated_size(self, *args, **kwargs):
+        """Returns the estimated size of the constraint.
+
+        Includes the size of the underlying expression (`Constraint.lhs`).
+
+        See [`Expression.estimated_size`][pyoframe.Expression.estimated_size] for details on signature and behavior.
+
+        Examples:
+            An dimensionless constraint has contains a 32 bit constraint id and, for each term, a 64 bit coefficient with a 32 bit variable id.
+            For a two-term expression that is: (32 + 2 * (64 + 32)) = 224 bits = 28 bytes.
+
+            >>> m = pf.Model()
+            >>> m.x = pf.Variable()
+            >>> m.con = m.x <= 4
+            >>> m.con.estimated_size()
+            28
+        """
+        return super().estimated_size(*args, **kwargs) + self.lhs.estimated_size(
+            *args, **kwargs
+        )
+
+    @overload
+    def to_str(self, return_df: Literal[False] = False) -> str: ...
+
+    @overload
+    def to_str(self, return_df: Literal[True] = True) -> pl.DataFrame: ...
+
+    def to_str(self, return_df: bool = False) -> str | pl.DataFrame:
+        """Converts the constraint to a human-readable string, or several arranged in a table.
+
+        Long expressions are truncated according to [`Config.print_max_terms`][pyoframe._Config.print_max_terms] and [`Config.print_polars_config`][pyoframe._Config.print_polars_config].
+
+        Parameters:
+            return_df:
+                If `True`, returns a DataFrame containing strings instead of the string representation of the DataFrame.
+
+        Examples:
+            >>> import polars as pl
+            >>> m = pf.Model()
+            >>> x = pf.Set(x=range(1000))
+            >>> y = pf.Set(y=range(1000))
+            >>> m.V = pf.Variable(x, y)
+            >>> expr = 2 * m.V * m.V
+            >>> print((expr <= 3).to_str())
+            ┌────────┬────────┬────────────────────────────────┐
+            │ x      ┆ y      ┆ constraint                     │
+            │ (1000) ┆ (1000) ┆                                │
+            ╞════════╪════════╪════════════════════════════════╡
+            │ 0      ┆ 0      ┆ 2 V[0,0] * V[0,0] <= 3         │
+            │ 0      ┆ 1      ┆ 2 V[0,1] * V[0,1] <= 3         │
+            │ 0      ┆ 2      ┆ 2 V[0,2] * V[0,2] <= 3         │
+            │ 0      ┆ 3      ┆ 2 V[0,3] * V[0,3] <= 3         │
+            │ 0      ┆ 4      ┆ 2 V[0,4] * V[0,4] <= 3         │
+            │ …      ┆ …      ┆ …                              │
+            │ 999    ┆ 995    ┆ 2 V[999,995] * V[999,995] <= 3 │
+            │ 999    ┆ 996    ┆ 2 V[999,996] * V[999,996] <= 3 │
+            │ 999    ┆ 997    ┆ 2 V[999,997] * V[999,997] <= 3 │
+            │ 999    ┆ 998    ┆ 2 V[999,998] * V[999,998] <= 3 │
+            │ 999    ┆ 999    ┆ 2 V[999,999] * V[999,999] <= 3 │
+            └────────┴────────┴────────────────────────────────┘
+            >>> expr = expr.sum("x")
+            >>> print((expr >= 3).to_str())
+            ┌────────┬─────────────────────────────────────────────────────────────────────────────────────────┐
+            │ y      ┆ constraint                                                                              │
+            │ (1000) ┆                                                                                         │
+            ╞════════╪═════════════════════════════════════════════════════════════════════════════════════════╡
+            │ 0      ┆ 2 V[0,0] * V[0,0] +2 V[1,0] * V[1,0] +2 V[2,0] * V[2,0] +2 V[3,0] * V[3,0] +2 V[4,0] *  │
+            │        ┆ V[4,0] … >= 3                                                                           │
+            │ 1      ┆ 2 V[0,1] * V[0,1] +2 V[1,1] * V[1,1] +2 V[2,1] * V[2,1] +2 V[3,1] * V[3,1] +2 V[4,1] *  │
+            │        ┆ V[4,1] … >= 3                                                                           │
+            │ 2      ┆ 2 V[0,2] * V[0,2] +2 V[1,2] * V[1,2] +2 V[2,2] * V[2,2] +2 V[3,2] * V[3,2] +2 V[4,2] *  │
+            │        ┆ V[4,2] … >= 3                                                                           │
+            │ 3      ┆ 2 V[0,3] * V[0,3] +2 V[1,3] * V[1,3] +2 V[2,3] * V[2,3] +2 V[3,3] * V[3,3] +2 V[4,3] *  │
+            │        ┆ V[4,3] … >= 3                                                                           │
+            │ 4      ┆ 2 V[0,4] * V[0,4] +2 V[1,4] * V[1,4] +2 V[2,4] * V[2,4] +2 V[3,4] * V[3,4] +2 V[4,4] *  │
+            │        ┆ V[4,4] … >= 3                                                                           │
+            │ …      ┆ …                                                                                       │
+            │ 995    ┆ 2 V[0,995] * V[0,995] +2 V[1,995] * V[1,995] +2 V[2,995] * V[2,995] +2 V[3,995] *       │
+            │        ┆ V[3,995] +2 V[4,99…                                                                     │
+            │ 996    ┆ 2 V[0,996] * V[0,996] +2 V[1,996] * V[1,996] +2 V[2,996] * V[2,996] +2 V[3,996] *       │
+            │        ┆ V[3,996] +2 V[4,99…                                                                     │
+            │ 997    ┆ 2 V[0,997] * V[0,997] +2 V[1,997] * V[1,997] +2 V[2,997] * V[2,997] +2 V[3,997] *       │
+            │        ┆ V[3,997] +2 V[4,99…                                                                     │
+            │ 998    ┆ 2 V[0,998] * V[0,998] +2 V[1,998] * V[1,998] +2 V[2,998] * V[2,998] +2 V[3,998] *       │
+            │        ┆ V[3,998] +2 V[4,99…                                                                     │
+            │ 999    ┆ 2 V[0,999] * V[0,999] +2 V[1,999] * V[1,999] +2 V[2,999] * V[2,999] +2 V[3,999] *       │
+            │        ┆ V[3,999] +2 V[4,99…                                                                     │
+            └────────┴─────────────────────────────────────────────────────────────────────────────────────────┘
+            >>> expr = expr.sum("y")
+            >>> print((expr == 3).to_str())
+            2 V[0,0] * V[0,0] +2 V[0,1] * V[0,1] +2 V[0,2] * V[0,2] +2 V[0,3] * V[0,3] +2 V[0,4] * V[0,4] … = 3
+        """
+        dims = self.dimensions
+        str_table = self.lhs.to_str(
+            include_const_term=False, return_df=True, str_col_name="constraint"
+        )
+        rhs = self.lhs.constant_terms.with_columns(pl.col(COEF_KEY) * -1)
+        rhs = cast_coef_to_string(rhs, drop_ones=False, always_show_sign=False)
+        rhs = rhs.rename({COEF_KEY: "rhs"})
+        if dims:
+            constr_str = str_table.join(
+                rhs, on=dims, how="left", maintain_order="left", coalesce=True
+            )
+        else:
+            constr_str = pl.concat([str_table, rhs], how="horizontal")
+        constr_str = constr_str.with_columns(
+            pl.concat_str("constraint", pl.lit(f" {self.sense.value} "), "rhs")
+        ).drop("rhs")
+
+        if not return_df:
+            if self.dimensions is None:
+                constr_str = constr_str.item()
+            else:
+                constr_str = self._add_shape_to_columns(constr_str)
+                with Config.print_polars_config:
+                    constr_str = repr(constr_str)
+
+        return constr_str
+
+    def __repr__(self) -> str:
+        return (
+            get_obj_repr(
+                self,
+                self.name,
+                height=len(self) if self.dimensions else None,
+                terms=len(self.lhs.data),
+                type=self.lhs.degree(return_str=True),
+            )
+            + "\n"
+            + self.to_str()
+        )
+
+
+class Variable(ModelElementWithId, SupportsMath):
+    """A decision variable for an optimization model.
+
+    Parameters:
+        *indexing_sets:
+            If no indexing_sets are provided, a single variable with no dimensions is created.
+            Otherwise, a variable is created for each element in the Cartesian product of the indexing_sets (see Set for details on behaviour).
+        lb:
+            The lower bound for all variables.
+        ub:
+            The upper bound for all variables.
+        vtype:
+            The type of the variable. Can be either a VType enum or a string. Default is VType.CONTINUOUS.
+        equals:
+            When specified, a variable is created and a constraint is added to make the variable equal to the provided expression.
+
+    Examples:
+        >>> import pandas as pd
+        >>> m = pf.Model()
+        >>> df = pd.DataFrame(
+        ...     {"dim1": [1, 1, 2, 2, 3, 3], "dim2": ["a", "b", "a", "b", "a", "b"]}
+        ... )
+        >>> Variable(df)
+        <Variable 'unnamed' height=6>
+        ┌──────┬──────┐
+        │ dim1 ┆ dim2 │
+        │ (3)  ┆ (2)  │
+        ╞══════╪══════╡
+        │ 1    ┆ a    │
+        │ 1    ┆ b    │
+        │ 2    ┆ a    │
+        │ 2    ┆ b    │
+        │ 3    ┆ a    │
+        │ 3    ┆ b    │
+        └──────┴──────┘
+
+        Variables cannot be used until they're added to the model.
+
+        >>> m.constraint = Variable(df) <= 3
+        Traceback (most recent call last):
+        ...
+        ValueError: Cannot use 'Variable' before it has been added to a model.
+
+        Instead, assign the variable to the model first:
+        >>> m.v = Variable(df)
+        >>> m.constraint = m.v <= 3
+
+        >>> m.v
+        <Variable 'v' height=6>
+        ┌──────┬──────┬──────────┐
+        │ dim1 ┆ dim2 ┆ variable │
+        │ (3)  ┆ (2)  ┆          │
+        ╞══════╪══════╪══════════╡
+        │ 1    ┆ a    ┆ v[1,a]   │
+        │ 1    ┆ b    ┆ v[1,b]   │
+        │ 2    ┆ a    ┆ v[2,a]   │
+        │ 2    ┆ b    ┆ v[2,b]   │
+        │ 3    ┆ a    ┆ v[3,a]   │
+        │ 3    ┆ b    ┆ v[3,b]   │
+        └──────┴──────┴──────────┘
+
+        >>> m.v2 = Variable(df[["dim1"]])
+        Traceback (most recent call last):
+        ...
+        ValueError: Duplicate rows found in input data.
+        >>> m.v3 = Variable(df[["dim1"]].drop_duplicates())
+        >>> m.v3
+        <Variable 'v3' height=3>
+        ┌──────┬──────────┐
+        │ dim1 ┆ variable │
+        │ (3)  ┆          │
+        ╞══════╪══════════╡
+        │ 1    ┆ v3[1]    │
+        │ 2    ┆ v3[2]    │
+        │ 3    ┆ v3[3]    │
+        └──────┴──────────┘
+    """
+
+    # TODO: Breaking change, remove support for Iterable[AcceptableSets]
+    def __init__(
+        self,
+        *indexing_sets: SetTypes | Iterable[SetTypes],
+        lb: float | int | SupportsToExpr | None = None,
+        ub: float | int | SupportsToExpr | None = None,
+        vtype: VType | VTypeValue = VType.CONTINUOUS,
+        equals: SupportsToExpr | None = None,
+    ):
+        if equals is not None:
+            assert len(indexing_sets) == 0, (
+                "Cannot specify both 'equals' and 'indexing_sets'"
+            )
+            indexing_sets = (equals,)
+
+        data = Set(*indexing_sets).data if len(indexing_sets) > 0 else pl.DataFrame()
+        super().__init__(data)
+
+        self.vtype: VType = VType(vtype)
+        self._attr = Container(self._set_attribute, self._get_attribute)
+        self._equals = equals
+
+        if lb is not None and not isinstance(lb, (float, int)):
+            self._lb_expr, self.lb = lb, None
+        else:
+            self._lb_expr, self.lb = None, lb
+        if ub is not None and not isinstance(ub, (float, int)):
+            self._ub_expr, self.ub = ub, None
+        else:
+            self._ub_expr, self.ub = None, ub
+
+    @property
+    def attr(self) -> Container:
+        """Allows reading and writing variable attributes similarly to [Model.attr][pyoframe.Model.attr]."""
+        return self._attr
+
+    def _set_attribute(self, name, value):
+        self._assert_has_ids()
+        col_name = name
+        try:
+            name = poi.VariableAttribute[name]
+            setter = self._model.poi.set_variable_attribute
+        except KeyError:
+            setter = self._model.poi.set_variable_raw_attribute
+
+        if self.dimensions is None:
+            for key in self.data.get_column(VAR_KEY):
+                setter(poi.VariableIndex(key), name, value)
+        else:
+            for key, v in (
+                self.data.join(
+                    value,
+                    on=self.dimensions,
+                    maintain_order="left" if Config.maintain_order else None,
+                )
+                .select(pl.col(VAR_KEY), pl.col(col_name))
+                .iter_rows()
+            ):
+                setter(poi.VariableIndex(key), name, v)
+
+    @unwrap_single_values
+    def _get_attribute(self, name):
+        self._assert_has_ids()
+        col_name = name
+        try:
+            name = poi.VariableAttribute[name]
+            getter = self._model.poi.get_variable_attribute
+        except KeyError:
+            getter = self._model.poi.get_variable_raw_attribute
+
+        ids = self.data.get_column(VAR_KEY).to_list()
+        attr = [getter(poi.VariableIndex(v_id), name) for v_id in ids]
+        data = self.data.with_columns(pl.Series(attr).alias(col_name))
+        return data.select(self._dimensions_unsafe + [col_name])
+
+    def _assign_ids(self):
+        assert self._model is not None
+        assert self.name is not None
+
+        solver = self._model.solver
+        if solver.supports_integer_variables:
+            domain = self.vtype._to_poi()
+        else:
+            if self.vtype != VType.CONTINUOUS:
+                raise ValueError(
+                    f"Solver {solver.name} does not support integer or binary variables."
+                )
+
+        lb = -1e100 if self.lb is None else float(self.lb)
+        ub = 1e100 if self.ub is None else float(self.ub)
+
+        poi_add_var = self._model.poi.add_variable
+
+        dims = self.dimensions
+
+        dynamic_names = dims is not None and self._model.solver_uses_variable_names
+        if dynamic_names:
+            names = concat_dimensions(self.data, prefix=self.name)[
+                "concated_dim"
+            ].to_list()
+            if solver.supports_integer_variables:
+                ids = [poi_add_var(domain, lb, ub, name).index for name in names]
+            else:
+                ids = [poi_add_var(lb, ub, name=name).index for name in names]
+        else:
+            if self._model.solver_uses_variable_names:
+                name = self.name
+            elif solver.block_auto_names:
+                name = "V"
+            else:
+                name = ""
+
+            n = 1 if dims is None else len(self.data)
+
+            if solver.supports_integer_variables:
+                ids = [poi_add_var(domain, lb, ub, name).index for _ in range(n)]
+            else:
+                ids = [poi_add_var(lb, ub, name=name).index for _ in range(n)]
+
+        df = self.data.with_columns(pl.Series(ids, dtype=KEY_TYPE).alias(VAR_KEY))
+
+        self._data = df
+
+    def _on_add_to_model(self, model, name):
+        super()._on_add_to_model(model, name)
+        self._assign_ids()
+        if self._lb_expr is not None:
+            setattr(model, f"{name}_lb", self._lb_expr <= self)
+
+        if self._ub_expr is not None:
+            setattr(model, f"{name}_ub", self <= self._ub_expr)
+
+        if self._equals is not None:
+            setattr(model, f"{name}_equals", self == self._equals)
+
+    @classmethod
+    def _get_id_column_name(cls):
+        return VAR_KEY
+
+    @property
+    @unwrap_single_values
+    def solution(self):
+        """Retrieves a variable's optimal value after the model has been solved.
+
+        Return type is a DataFrame if the variable has dimensions, otherwise it is a single value.
+        Binary and integer variables are returned as integers.
+
+        Examples:
+            >>> m = pf.Model()
+            >>> m.var_continuous = pf.Variable({"dim1": [1, 2, 3]}, lb=5, ub=5)
+            >>> m.var_integer = pf.Variable(
+            ...     {"dim1": [1, 2, 3]}, lb=4.5, ub=5.5, vtype=pf.VType.INTEGER
+            ... )
+            >>> m.var_dimensionless = pf.Variable(
+            ...     lb=4.5, ub=5.5, vtype=pf.VType.INTEGER
+            ... )
+            >>> m.var_continuous.solution
+            Traceback (most recent call last):
+            ...
+            RuntimeError: Failed to retrieve solution for variable. Are you sure the model has been solved?
+            >>> m.optimize()
+            >>> m.var_continuous.solution
+            shape: (3, 2)
+            ┌──────┬──────────┐
+            │ dim1 ┆ solution │
+            │ ---  ┆ ---      │
+            │ i64  ┆ f64      │
+            ╞══════╪══════════╡
+            │ 1    ┆ 5.0      │
+            │ 2    ┆ 5.0      │
+            │ 3    ┆ 5.0      │
+            └──────┴──────────┘
+            >>> m.var_integer.solution
+            shape: (3, 2)
+            ┌──────┬──────────┐
+            │ dim1 ┆ solution │
+            │ ---  ┆ ---      │
+            │ i64  ┆ i64      │
+            ╞══════╪══════════╡
+            │ 1    ┆ 5        │
+            │ 2    ┆ 5        │
+            │ 3    ┆ 5        │
+            └──────┴──────────┘
+            >>> m.var_dimensionless.solution
+            5
+        """
+        try:
+            solution = self.attr.Value
+        except RuntimeError as e:
+            raise RuntimeError(
+                "Failed to retrieve solution for variable. Are you sure the model has been solved?"
+            ) from e
+        if isinstance(solution, pl.DataFrame):
+            solution = solution.rename({"Value": SOLUTION_KEY})
+
+        if self.vtype in [VType.BINARY, VType.INTEGER]:
+            if isinstance(solution, pl.DataFrame):
+                # TODO handle values that are out of bounds of Int64 (i.e. when problem is unbounded)
+                solution = solution.with_columns(
+                    pl.col("solution").alias("solution_float"),
+                    pl.col("solution").round().cast(pl.Int64),
+                )
+                if Config.integer_tolerance != 0:
+                    df = solution.filter(
+                        (pl.col("solution_float") - pl.col("solution")).abs()
+                        > Config.integer_tolerance
+                    )
+                    assert df.is_empty(), (
+                        f"Variable {self.name} has a non-integer value: {df}\nThis should not happen."
+                    )
+                solution = solution.drop("solution_float")
+            else:
+                solution_float = solution
+                solution = int(round(solution))
+                if Config.integer_tolerance != 0:
+                    assert abs(solution - solution_float) < Config.integer_tolerance, (
+                        f"Value of variable {self.name} is not an integer: {solution}. This should not happen."
+                    )
+
+        return solution
+
+    def __repr__(self):
+        result = (
+            get_obj_repr(
+                self,
+                self.name,
+                lb=self.lb,
+                ub=self.ub,
+                height=self.data.height if self.dimensions else None,
+            )
+            + "\n"
+        )
+        if self._has_ids:
+            result += self.to_expr().to_str(str_col_name="variable")
+        else:
+            with Config.print_polars_config:
+                data = self._add_shape_to_columns(self.data)
+                # we don't try to include the allowed_new_dims because there are none for Variables (only exist on Expression or Sets)
+                result += repr(data)
+
+        return result
+
+    def to_expr(self) -> Expression:
+        """Converts the Variable to an Expression."""
+        self._assert_has_ids()
+        return self._new(self.data.drop(SOLUTION_KEY, strict=False), self.name)  # pyright: ignore[reportArgumentType], we know it's safe after _assert_has_ids()
+
+    def _new(self, data: pl.DataFrame, name: str) -> Expression:
+        self._assert_has_ids()
+        e = Expression(data.with_columns(pl.lit(1.0).alias(COEF_KEY)), name)
+        e._model = self._model
+        return e
+
+    @return_new
+    def next(self, dim: str, wrap_around: bool = False):
+        """Creates an expression where the variable at each index is the next variable in the specified dimension.
+
+        Parameters:
+            dim:
+                The dimension over which to shift the variable.
+            wrap_around:
+                If `True`, the last index in the dimension is connected to the first index.
+
+        Examples:
+            >>> import pandas as pd
+            >>> time_dim = pd.DataFrame({"time": ["00:00", "06:00", "12:00", "18:00"]})
+            >>> space_dim = pd.DataFrame({"city": ["Toronto", "Berlin"]})
+            >>> m = pf.Model()
+            >>> m.bat_charge = pf.Variable(time_dim, space_dim)
+            >>> m.bat_flow = pf.Variable(time_dim, space_dim)
+            >>> # Fails because the dimensions are not the same
+            >>> m.bat_charge + m.bat_flow == m.bat_charge.next("time")
+            Traceback (most recent call last):
+            ...
+            pyoframe._constants.PyoframeError: Cannot subtract the two expressions below because of unmatched values.
+            Expression 1:   (bat_charge + bat_flow)
+            Expression 2:   bat_charge.next(…)
+            Unmatched values:
+            shape: (2, 4)
+            ┌───────┬─────────┬────────────┬────────────┐
+            │ time  ┆ city    ┆ time_right ┆ city_right │
+            │ ---   ┆ ---     ┆ ---        ┆ ---        │
+            │ str   ┆ str     ┆ str        ┆ str        │
+            ╞═══════╪═════════╪════════════╪════════════╡
+            │ 18:00 ┆ Toronto ┆ null       ┆ null       │
+            │ 18:00 ┆ Berlin  ┆ null       ┆ null       │
+            └───────┴─────────┴────────────┴────────────┘
+            If this is intentional, use .drop_unmatched() or .keep_unmatched().
+
+            >>> (m.bat_charge + m.bat_flow).drop_unmatched() == m.bat_charge.next(
+            ...     "time"
+            ... )
+            <Constraint 'unnamed' height=6 terms=18 type=linear>
+            ┌───────┬─────────┬────────────────────────────────────────────────────────────────────────────────┐
+            │ time  ┆ city    ┆ constraint                                                                     │
+            │ (3)   ┆ (2)     ┆                                                                                │
+            ╞═══════╪═════════╪════════════════════════════════════════════════════════════════════════════════╡
+            │ 00:00 ┆ Toronto ┆ bat_charge[00:00,Toronto] + bat_flow[00:00,Toronto]                            │
+            │       ┆         ┆ - bat_charge[06:00,Toronto] = 0                                                │
+            │ 00:00 ┆ Berlin  ┆ bat_charge[00:00,Berlin] + bat_flow[00:00,Berlin] - bat_charge[06:00,Berlin]   │
+            │       ┆         ┆ = 0                                                                            │
+            │ 06:00 ┆ Toronto ┆ bat_charge[06:00,Toronto] + bat_flow[06:00,Toronto]                            │
+            │       ┆         ┆ - bat_charge[12:00,Toronto] = 0                                                │
+            │ 06:00 ┆ Berlin  ┆ bat_charge[06:00,Berlin] + bat_flow[06:00,Berlin] - bat_charge[12:00,Berlin]   │
+            │       ┆         ┆ = 0                                                                            │
+            │ 12:00 ┆ Toronto ┆ bat_charge[12:00,Toronto] + bat_flow[12:00,Toronto]                            │
+            │       ┆         ┆ - bat_charge[18:00,Toronto] = 0                                                │
+            │ 12:00 ┆ Berlin  ┆ bat_charge[12:00,Berlin] + bat_flow[12:00,Berlin] - bat_charge[18:00,Berlin]   │
+            │       ┆         ┆ = 0                                                                            │
+            └───────┴─────────┴────────────────────────────────────────────────────────────────────────────────┘
+
+            >>> (m.bat_charge + m.bat_flow) == m.bat_charge.next(
+            ...     "time", wrap_around=True
+            ... )
+            <Constraint 'unnamed' height=8 terms=24 type=linear>
+            ┌───────┬─────────┬────────────────────────────────────────────────────────────────────────────────┐
+            │ time  ┆ city    ┆ constraint                                                                     │
+            │ (4)   ┆ (2)     ┆                                                                                │
+            ╞═══════╪═════════╪════════════════════════════════════════════════════════════════════════════════╡
+            │ 00:00 ┆ Toronto ┆ bat_charge[00:00,Toronto] + bat_flow[00:00,Toronto]                            │
+            │       ┆         ┆ - bat_charge[06:00,Toronto] = 0                                                │
+            │ 00:00 ┆ Berlin  ┆ bat_charge[00:00,Berlin] + bat_flow[00:00,Berlin] - bat_charge[06:00,Berlin]   │
+            │       ┆         ┆ = 0                                                                            │
+            │ 06:00 ┆ Toronto ┆ bat_charge[06:00,Toronto] + bat_flow[06:00,Toronto]                            │
+            │       ┆         ┆ - bat_charge[12:00,Toronto] = 0                                                │
+            │ 06:00 ┆ Berlin  ┆ bat_charge[06:00,Berlin] + bat_flow[06:00,Berlin] - bat_charge[12:00,Berlin]   │
+            │       ┆         ┆ = 0                                                                            │
+            │ 12:00 ┆ Toronto ┆ bat_charge[12:00,Toronto] + bat_flow[12:00,Toronto]                            │
+            │       ┆         ┆ - bat_charge[18:00,Toronto] = 0                                                │
+            │ 12:00 ┆ Berlin  ┆ bat_charge[12:00,Berlin] + bat_flow[12:00,Berlin] - bat_charge[18:00,Berlin]   │
+            │       ┆         ┆ = 0                                                                            │
+            │ 18:00 ┆ Toronto ┆ bat_charge[18:00,Toronto] + bat_flow[18:00,Toronto]                            │
+            │       ┆         ┆ - bat_charge[00:00,Toronto] = 0                                                │
+            │ 18:00 ┆ Berlin  ┆ bat_charge[18:00,Berlin] + bat_flow[18:00,Berlin] - bat_charge[00:00,Berlin]   │
+            │       ┆         ┆ = 0                                                                            │
+            └───────┴─────────┴────────────────────────────────────────────────────────────────────────────────┘
+
+        """
+        wrapped = (
+            self.data.select(dim)
+            .unique(maintain_order=Config.maintain_order)
+            .sort(by=dim)
+        )
+        wrapped = wrapped.with_columns(pl.col(dim).shift(-1).alias("__next"))
+        if wrap_around:
+            wrapped = wrapped.with_columns(pl.col("__next").fill_null(pl.first(dim)))
+        else:
+            wrapped = wrapped.drop_nulls(dim)
+
+        expr = self.to_expr()
+        data = expr.data.rename({dim: "__prev"})
+
+        data = data.join(
+            wrapped,
+            left_on="__prev",
+            right_on="__next",
+            # We use "right" instead of "left" to maintain consistency with the behavior without maintain_order
+            maintain_order="right" if Config.maintain_order else None,
+        ).drop(["__prev", "__next"], strict=False)
+
+        return data