IEEE 754 Double-Precision Standard

The theoretical floating-point system we have been studying becomes real hardware through the IEEE 754 standard. The 64-bit double-precision format is the workhorse of scientific computing.

---

64-Bit Layout

A double-precision number occupies exactly 64 bits, split into three fields:

Field	Bits	Purpose
Sign	1	0 = positive, 1 = negative
Exponent	11	Stores the (biased) exponent
Fraction	52	Stores $d_1 d_2 \ldots d_{52}$ of the mantissa

---

The Denormalized Form

IEEE 754 uses Convention 3 (Denormalized Form). Every normal number is interpreted as:

$\pm\,(1.d_1\,d_2\,\ldots\,d_{52})_2 \;\times\; 2^{\,e}$

The leading 1 before the decimal point is always present for normal numbers and is implicit. It is not stored in the 52 fraction bits, giving one extra bit of precision for free.

---

Why Exponent Biasing Is Necessary

The 11 exponent bits can hold unsigned integers from 0 to $2^{11}-1 = 2047$.
Without biasing, the smallest representable number would be:

$(1.00\ldots 0)_2 \times 2^0 = 1$

That is far too large since it cannot represent numbers like $0.001$. Biasing fixes this by interpreting the stored value as $e_{\text{effective}} = e_{\text{stored}} - \text{bias}$.

---

Exponent Biasing (Bias = 1023)

We subtract 1023 from the stored exponent to get the effective exponent. So, in denormalized form, we can write it as:

$\pm\,(1.d_1\,d_2\,\ldots\,d_{52})_2 \;\times\; 2^{\,e - 1023}$

Using the equivalent 0.1 form:

$\pm\,(0.1\,d_1\,d_2\,\ldots\,d_{52})_2 \;\times\; 2^{\,e - 1023}\times 2^1$ or, $\pm\,(0.1\,d_1\,d_2\,\ldots\,d_{52})_2 \;\times\; 2^{\,e - 1022}$

Naive Minimum and Maximum

Naively, if we used every possible stored exponent value, our minimum and maximum would be:

• Naive Minimum: Storing $0$ yields an effective exponent of $-1022$ $(0.100\ldots 0)_2 \times 2^{-1022}$
• Naive Maximum: Storing $2047$ yields an effective exponent of $1025$ $(0.111\ldots 1)_2 \times 2^{1025}$

---

The Problem: How Do We Represent Zero?

There is a major flaw with the naive approach above: we cannot represent exactly zero.

Because denormalized numbers rely on an implicit leading 1 ($1.d_1d_2\dots$), the fraction part is always at least $1.0$. No matter how small the exponent gets, the final calculated value never truly reaches $0$.

---

The Solution: Reserved Exponents

To fix the zero problem (and handle math errors), IEEE 754 sacrifices the two extreme stored exponents ($0$ and $2047$) and reserves them for special cases:

Stored exponent	Fraction bits	Represents
0	all zeros	$\pm 0$
0	any non-zero	very small numbers near 0, consider as 0
2047	all zeros	$\pm\infty$ (overflow result)
2047	any non-zero	Not a Number (e.g. $0/0$)

---

Overflow and Underflow

With the extremes reserved, pushing outside the boundaries of normal numbers triggers specific behaviors:

Condition	When	Result
Overflow	Computed exponent $> 1024$	Stored as $\pm\infty$
Underflow	Computed exponent $< -1021$	Stored as 0

---

Actual Practical Limits

By reserving $0$ and $2047$, the usable stored exponents for are strictly bounded between 1 and 2046. This gives us an effective exponent range of $-1021$ to $1024$.

Here are the actual limits for double-precision normal numbers:

$\text{Smallest normal} \approx 2.225 \times 10^{-308}$

$\text{Largest normal} \approx 1.798 \times 10^{308}$

These are the limits you see when a language reports Double.MAX_VALUE or DBL_MAX.

Where Do These Come From?

Smallest normal: Using the minimum usable exponent ($1 - 1022 = -1021$) and all zeros in the fraction:

$(0.100\ldots 0)_2 \times 2^{-1021} = 2^{-1021} \approx 2.225 \times 10^{-308}$

Largest normal: Using the maximum usable exponent ($2046 - 1022 = 1024$) and all ones in the fraction:

$(0.111\ldots 1)_2 \times 2^{1024} \approx (2 - 2^{-52}) \times 2^{1024} \approx 1.798 \times 10^{308}$

---

Caution

Key Takeaway: IEEE 754 double precision ($\beta=2$, 52 mantissa bits, 11 exponent bits, bias 1023) provides about 15–17 significant decimal digits of precision and a range spanning roughly $10^{-308}$ to $10^{308}$. Two exponent bit patterns are reserved for zero/subnormals and infinity/NaN, enabling graceful handling of exceptional conditions.

---